зеркало из https://github.com/mozilla/kaldi.git
trunk: copying from sandbox/oplatek2, the vystadial_en and vystadial_cz recipes
git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3809 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8
This commit is contained in:
Dan Povey
Родитель
01c4041da7
Коммит
528345f30c
|
|
@ -0,0 +1,168 @@
|
|||
Summary
|
||||
-------
|
||||
The data comprise over 15 hours of speech in Czech.
|
||||
|
||||
The Czech recordings were collected in three ways:
|
||||
|
||||
1. using a free Call Friend phone service
|
||||
2. using the Repeat After Me speech data collecting process
|
||||
3. from telephone interactions with the PublicTransportInfo Spoken Dialog System (SDS)
|
||||
Alex: http://ufal.ms.mff.cuni.cz/alex-dialogue-systems-framework/.
|
||||
|
||||
The data collection process is described in detail
|
||||
in article "Free English and Czech telephone speech corpus shared under the CC-BY-SA 3.0 license"
|
||||
published for LREC 2014 (To Appear).
|
||||
|
||||
WE USE COMMON KALDI DECODERS IN THE SCRIPTS (gmm-latgen-faster through steps/decode.sh)
|
||||
However, the main purpose of providing the data and scripts
|
||||
is training acoustic models for real-time speech recognition unit
|
||||
for dialog system ALEX, which uses modified real-time Kaldi OnlineLatgenRecogniser.
|
||||
The modified Kaldi decoders are NOT required for running the scripts!
|
||||
|
||||
The modified OnlineLatgenRecogniser is actively developed at
|
||||
https://github.com/UFAL-DSG/pykaldi/tree/master/src/onl-rec
|
||||
and has Python wrapper:
|
||||
https://github.com/UFAL-DSG/pykaldi/tree/master/src/pykaldi
|
||||
Note that I am currently moving the online recogniser to:
|
||||
http://sourceforge.net/p/kaldi/code/HEAD/tree/sandbox/oplatek2/
|
||||
|
||||
Credits and license
|
||||
------------------------
|
||||
The scripts are partially based on Voxforge KALDI recipe.
|
||||
The original scripts as well as theses scripts are licensed under APACHE 2.0 license.
|
||||
The data are distributed under Attribution-{ShareAlike} 3.0 Unported ({CC} {BY}-{SA} 3.0) license.
|
||||
Czech data: https://lindat.mff.cuni.cz/repository/xmlui/handle/11858/00-097C-0000-0023-4670-6
|
||||
English data: https://lindat.mff.cuni.cz/repository/xmlui/handle/11858/00-097C-0000-0023-4671-4
|
||||
|
||||
The data collecting process and development of these training scripts
|
||||
was partly funded by the Ministry of Education, Youth and Sports
|
||||
of the Czech Republic under the grant agreement LK11221
|
||||
and core research funding of Charles University in Prague.
|
||||
For citing, please use following BibTex citation:
|
||||
|
||||
@inproceedings{korvas_2014,
|
||||
title={{Free English and Czech telephone speech corpus shared under the CC-BY-SA 3.0 license}},
|
||||
author={Korvas, Mat\v{e}j and Pl\'{a}tek, Ond\v{r}ej and Du\v{s}ek, Ond\v{r}ej and \v{Z}ilka, Luk\'{a}\v{s} and Jur\v{c}\'{i}\v{c}ek, Filip},
|
||||
booktitle={Proceedings of the Eigth International Conference on Language Resources and Evaluation (LREC 2014)},
|
||||
pages={To Appear},
|
||||
year={2014},
|
||||
}
|
||||
|
||||
|
||||
Expected results
|
||||
----------------
|
||||
The expected results were obtained simply by running
|
||||
bash train_voip_cs.sh OR bash train_voip_en.sh.
|
||||
Note that you need SRILM installed in path or at kaldi/tools/ directory!
|
||||
|
||||
build2 - bigram LM from train data, estimated by the scripts using SRILM
|
||||
build0 - zerogram LM from test data, estimated by scripts using Python code
|
||||
LMW - Language model weight, we picked the best from (min_lmw, max_lmw) based on decoding results on DEV set
|
||||
|
||||
Full Czech data:
|
||||
exp set LM LMW WER SER
|
||||
mono test build0 6 86.1 89.66
|
||||
tri1 test build0 8 70.84 82.9
|
||||
tri2a test build0 8 70.86 83.01
|
||||
tri2b test build0 9 68.13 80.89
|
||||
tri2b_mmi test build0 9 67.61 79.53
|
||||
tri2b_mmi_b0.05 test build0 8 66.18 78.72
|
||||
tri2b_mpe test build0 9 64.93 77.66
|
||||
mono test build2 8 72.3 79.02
|
||||
tri1 test build2 11 55.57 72.11
|
||||
tri2a test build2 11 55.12 70.9
|
||||
tri2b test build2 12 52.95 70.7
|
||||
tri2b_mmi test build2 10 50.42 68.38
|
||||
tri2b_mmi_b0.05 test build2 10 49.96 68.58
|
||||
tri2b_mpe test build2 12 49.87 66.97
|
||||
|
||||
Note that the zero-gram LMs for discriminative training
|
||||
give significant advantage, because they are estimated on test set!
|
||||
|
||||
|
||||
Details
|
||||
-------
|
||||
* Requires Kaldi installation and Linux environment. (Tested on Ubuntu 10.04, 12.04 and 12.10.)
|
||||
* The config files s5/env_voip_cs.sh sets the data directory,
|
||||
mfcc directory and experiments directory.
|
||||
* Our scripts prepare the data to the expected format in s5/data.
|
||||
* Experiment files are stored to $exp directory e.g. s5/exp.
|
||||
* The local directory contains scripts for data preparation to prepare
|
||||
lang directory.
|
||||
* path.sh, cmd.sh and common/* contain configurations for the
|
||||
recipe.
|
||||
* Language model (LM) is either built from the training data using
|
||||
[SRILM](http://www.speech.sri.com/projects/srilm/) or we supply one in
|
||||
the ARPA format.
|
||||
|
||||
|
||||
Running experiments
|
||||
-------------------
|
||||
Before running the experiments, check that:
|
||||
|
||||
* you have the Kaldi toolkit compiled:
|
||||
http://sourceforge.net/projects/kaldi/.
|
||||
* you have SRILM compiled. (This is needed for building a language model
|
||||
unless you supply your own LM in the ARPA format.)
|
||||
See http://www.speech.sri.com/projects/srilm/.
|
||||
* The number of jobs njobs are set correctly in path.sh.
|
||||
* In cmd.sh, you switched to run the training on a SGE[*] grid if
|
||||
required (disabled by default).
|
||||
|
||||
Start the recipe from the s5 directory by running
|
||||
bash run.sh.
|
||||
It will create s5/mfcc, s5/data and s5/exp directories.
|
||||
If any of them exists, it will ask you if you want them to be overwritten.
|
||||
|
||||
.. [*] Sun Grid Engine
|
||||
|
||||
Extracting the results and trained models
|
||||
-----------------------------------------
|
||||
The main scripts, s5/run.sh,
|
||||
perform not only training of the acoustic models, but also decoding.
|
||||
The acoustic models are evaluated after running the training and
|
||||
reports are printed to the standard output.
|
||||
|
||||
The s5/local/results.py exp command extracts the results from the $exp directory.
|
||||
and stores the results to exp/results.log.
|
||||
|
||||
If you want to use the trained acoustic model with your language model
|
||||
outside the prepared script, you need to build the HCLG decoding graph yourself.
|
||||
See http://kaldi.sourceforge.net/graph.html for general introduction to the FST
|
||||
framework in Kaldi.
|
||||
|
||||
The simplest way to start decoding is to use the same LM which
|
||||
was used by the s5/run.sh script.
|
||||
Let's say you want to decode with
|
||||
the acoustic model stored in exp/tri2b_bmmi,
|
||||
then you need files listed below:
|
||||
|
||||
================================= =====================================================================================
|
||||
mfcc.conf Speech parametrisation (MFCC) settings. Training and decoding setup must match.
|
||||
exp/tri2b_bmmi/graph/HCLG.fst Decoding Graph. Graph part of AM plus lexicon, phone->3phone & LM representation.
|
||||
exp/tri2b_bmmi/graph/words.txt Word symbol table, a mapping between words and integers which are decoded.
|
||||
exp/tri2b_bmmi/graph/silence.csl List of phone integer ids, which represent silent phones.
|
||||
exp/tri2b_bmmi/final.mdl Trained acoustic model (AM).
|
||||
exp/tri2b_bmmi/final.mat Trained matrix of feature/space transformations (E.g. LDA and bMMI).
|
||||
================================= =====================================================================================
|
||||
|
||||
|
||||
We recommend to study steps/decode.sh Kaldi standard script
|
||||
for standalone decoding with gmm-latgen-faster Kaldi decoder.
|
||||
|
||||
In order to build your own decoding graph HCLG
|
||||
you need LM in ARPA format and files in table below.
|
||||
|
||||
* Note 1: Building HCLG decoding graph is out of scope this README.
|
||||
* Note 2: Each acoustic model needs corresponding HCLG graph.
|
||||
* Note 3: The phonetic dictionary applied on the vocabulary
|
||||
should always generate only a subset of phones seen in training data!
|
||||
|
||||
=============================== =========================================================================
|
||||
LM.arpa Language model in ARPA format [You should supply it]
|
||||
vocabulary.txt List of words you want to decode [You should supply it]
|
||||
OOV_SYMBOL String representing out of vocabulary word. [You should supply it]
|
||||
dictionary.txt Phonetic dictionary. [You should supply it]
|
||||
exp/tri2b_bmmi/final.mdl Trained acoustic model (AM).
|
||||
exp/tri2b_bmmi/final.tree Phonetic decision tree.
|
||||
=============================== =========================================================================
|
||||
|
|
@ -0,0 +1,37 @@
|
|||
###############################
|
||||
# Examples commands for gdb #
|
||||
###############################
|
||||
# layout src # Open TUI mode in viewing src code
|
||||
# file ./online-python-gmm-decode-faster-test # we want to debug this program
|
||||
# directory ../feat # add directory to search path for source codes
|
||||
# b online-python-gmm-decode-faster-test.cc:80 # set breakpoint
|
||||
# set args --rt-min=0.5 --rt-max=0.7 # pass arguments to program
|
||||
# shell ls # execute shell command ls
|
||||
|
||||
#############################
|
||||
# EXAMPLES how to run gdb #
|
||||
#############################
|
||||
# gdb -q -iex "set auto-load safe-path ." .gdbinit
|
||||
# gdb -q -x .gdbinit
|
||||
|
||||
#######################
|
||||
# USEFULL shortcuts #
|
||||
#######################
|
||||
# Ctrl+x Ctrl+a ... switches of and of tui from gdb prompt
|
||||
# In TUI mode: Ctrl+n resp. Ctrl+p ... next resp. previous line in history
|
||||
|
||||
###########
|
||||
# LINKS #
|
||||
###########
|
||||
# Martin Jiricka slides in czech:
|
||||
# http://www.ms.mff.cuni.cz/~jirim7am/data/gdb/gdb.pdf (or in my Calibre library)
|
||||
|
||||
directory ../../../src/onl-rec
|
||||
directory ../../../src/decoder
|
||||
directory ../../../src/feat
|
||||
|
||||
# extractor->Compute
|
||||
b onl-rec-audio-source.cc:49
|
||||
b onl-rec-audio-source.cc:31
|
||||
|
||||
run
|
||||
|
|
@ -0,0 +1,64 @@
|
|||
BEST_LINE=18
|
||||
MODEL_PREFIX_URL=http://vystadial.ms.mff.cuni.cz/download/kaldi/src/pykaldi/pykaldi/binutils/
|
||||
DATA_PREFIX_URL=http://vystadial.ms.mff.cuni.cz/download/kaldi/src/pykaldi/pykaldi/binutils/
|
||||
|
||||
# Czech language models
|
||||
LANG=cs
|
||||
HCLG=models/HCLG_tri2b_bmmi.fst
|
||||
AM=models/tri2b_bmmi.mdl
|
||||
MAT=models/tri2b_bmmi.mat # matrix trained in tri2b models
|
||||
WST=models/words.txt
|
||||
MFCC=models/mfcc.conf
|
||||
SILENCE=models/silence.csl
|
||||
|
||||
$(MFCC):
|
||||
wget $(MODEL_PREFIX_URL)/$@ -O $@
|
||||
|
||||
$(SILENCE):
|
||||
wget $(MODEL_PREFIX_URL)/$@ -O $@
|
||||
|
||||
$(HCLG):
|
||||
wget $(MODEL_PREFIX_URL)/$@ -O $@
|
||||
|
||||
$(AM):
|
||||
wget $(MODEL_PREFIX_URL)/$@ -O $@
|
||||
|
||||
$(MAT):
|
||||
wget $(MODEL_PREFIX_URL)/$@ -O $@
|
||||
|
||||
$(WST):
|
||||
wget $(MODEL_PREFIX_URL)/$@ -O $@
|
||||
|
||||
data/vystadial-sample-$(LANG).tar.gz:
|
||||
wget $(DATA_PREFIX_URL)/$@ -O $@
|
||||
|
||||
data/vystadial-sample-$(LANG)/README.rst: data/vystadial-sample-$(LANG).tar.gz
|
||||
tar xf $< -C data
|
||||
touch $@
|
||||
|
||||
data/vystadial-sample-$(LANG)/test/input.scp: data/vystadial-sample-$(LANG)/README.rst
|
||||
for f in `dirname $@`/*.wav ; do echo "`basename $$f`" "`pwd`/$$f" ; done > $@
|
||||
|
||||
data/vystadial-sample-$(LANG)/test/input_best.scp: data/vystadial-sample-$(LANG)/test/input.scp
|
||||
sed -n '$(BEST_LINE)p' < $< > $@
|
||||
|
||||
build_scp: data/vystadial-sample-$(LANG)/test/input_best.scp data/vystadial-sample-$(LANG)/test/input.scp
|
||||
|
||||
download_models: $(SILENCE) $(WST) $(HCLG) $(AM) $(MAT) $(MFCC)
|
||||
|
||||
pyonline-recogniser: build_scp download_models
|
||||
bash run_pyonline-latgen-recogniser.sh
|
||||
|
||||
online-recogniser: build_scp download_models
|
||||
bash run_online-latgen-recogniser.sh
|
||||
|
||||
gmm-latgen-faster: build_scp download_models
|
||||
bash run_gmm-latgen-faster.sh
|
||||
|
||||
live: download_models
|
||||
bash run_live-demo.sh
|
||||
|
||||
clean:
|
||||
rm -rf *.pyc data/* models/* decode/*
|
||||
|
||||
.PHONY: live pyonline-recogniser online-recogniser gmm-latgen-faster clean
|
||||
|
|
@ -0,0 +1,80 @@
|
|||
Running the example Pykaldi scripts
|
||||
===================================
|
||||
|
||||
Summary
|
||||
-------
|
||||
The demo presents three new Kaldi features on pretrained Czech AMs:
|
||||
* Online Lattice Recogniser. The best results were obtained using MFCC, LDA+MLLT and bMMI.
|
||||
* Python wrapper which interfaces the OnlineLatticeRecogniser to Python.
|
||||
* Training scripts which can be used with standard Kaldi tools or with the new OnlineLatticeRecogniser.
|
||||
|
||||
The pykaldi-latgen-faster-decoder.py
|
||||
demonstrates how to use the class PyOnlineLatgenRecogniser,
|
||||
which takes audio on the input and outputs the decoded lattice.
|
||||
There are also the OnlineLatgenRecogniser C++ and Kaldi standard gmm-latgen-faster demos.
|
||||
All three demos produce the same results.
|
||||
|
||||
TODO: Publish English AM and add English demo
|
||||
|
||||
In March 2014, the PyOnlineLatticeRecogniser recogniser was evaluated on domain of SDS Alex.
|
||||
See graphs evaluating OnlineLatticeRecogniser performance at
|
||||
http://nbviewer.ipython.org/github/oplatek/pykaldi-eval/blob/master/Pykaldi-evaluation.ipynb.
|
||||
|
||||
An example posterior word lattice output for one Czech utterance can be seen at
|
||||
http://oplatek.blogspot.it/2014/02/ipython-demo-pykaldi-decoders-on-short.html
|
||||
|
||||
Dependencies
|
||||
------------
|
||||
* Build (make) and test (make test) the code under kaldi/src, kaldi/src/pykaldi and kaldi/src/onl-rec
|
||||
* For inspecting the saved lattices you need dot binary
|
||||
from Graphviz <http://www.graphviz.org/Download..php library.
|
||||
* For running the live demo you need pyaudio package.
|
||||
|
||||
Running the example scripts
|
||||
---------------------------
|
||||
|
||||
|
||||
make online-latgen-recogniser
|
||||
|
||||
* Run the test src/onl-rec/onl-rec-latgen-recogniser-test for OnlineLatgenRecogniser
|
||||
which shows C++ example of how to use the recogniser.
|
||||
The same data, AM a LM are used as for make pyonline-latgen-recogniser.
|
||||
The pretrained Language (LM) and Acoustic (AM) models are used.
|
||||
The data as well as the models are downloaded from our server.
|
||||
|
||||
|
||||
make pyonline-latgen-recogniser
|
||||
|
||||
* Run the decoding with PyOnlineFasterRecogniser.
|
||||
Example Python script pykaldi-online-latgen-recogniser.py shows
|
||||
PyOnlineFasterRecogniser decoding on small test set.
|
||||
The same pretrained Language (LM) and Acoustic (AM) models.
|
||||
|
||||
|
||||
make gmm-latgen-faster
|
||||
|
||||
* Run the decoding with Kaldi gmm-latgen-faster executable wrapped in `<run_gmm-latgen-faster.sh>`_.
|
||||
This is the reference executable for
|
||||
The same data, AM a LM are used as for make pyonline-latgen-recogniser.
|
||||
We use this script as reference.
|
||||
|
||||
|
||||
make live
|
||||
|
||||
* The simple live demo should decode speech from your microphone.
|
||||
It uses the pretrained AM and LM and wraps `<live-demo.py>`_.
|
||||
The pyaudio package is used for capturing the sound from your microphone.
|
||||
We were able to use it under `Ubuntu 12.10` and Python 2.7, but we guarantee nothing on your system.
|
||||
|
||||
|
||||
Notes
|
||||
-----
|
||||
The scripts for Czech and English support acoustic models obtained using MFCC, LDA+MLLT/delta+delta-delta feature transformations and acoustic models trained generatively or by MPE or bMMI training.
|
||||
|
||||
The new functionality is separated to different directories:
|
||||
* kaldi/src/onl-rec stores C++ code for OnlineLatticeRecogniser.
|
||||
* kaldi/scr/pykaldi stores Python wrapper PyOnlineLatticeRecogniser.
|
||||
* kaldi/egs/vystadial/s5 stores training scripts.
|
||||
* kaldi/egs/vystadial/online_demo shows Kaldi standard decoder, OnlineLatticeRecogniser and PyOnlineLatticeRecogniser, which produce the exact same lattices using the same setup.
|
||||
|
||||
The OnlineLatticeRecogniser is used in Alex dialogue system (https://github.com/UFAL-DSG/alex).
|
||||
|
|
@ -0,0 +1 @@
|
|||
from __future__ import unicode_literals
|
||||
|
|
@ -0,0 +1,54 @@
|
|||
#!/usr/bin/env python
|
||||
# encoding: utf-8
|
||||
from __future__ import unicode_literals
|
||||
|
||||
import glob
|
||||
import sys
|
||||
import os
|
||||
import codecs
|
||||
|
||||
def build_reference(wav_scp, ref_path):
|
||||
print wav_scp, ref_path
|
||||
with codecs.open(ref_path, 'w', 'utf-8') as w:
|
||||
with codecs.open(wav_scp, 'r', 'utf-8') as scp:
|
||||
for line in scp:
|
||||
name, wavpath = line.strip().split(' ', 1)
|
||||
with codecs.open(wavpath + '.trn', 'r', 'utf-8') as trn:
|
||||
trans = trn.read().strip()
|
||||
w.write(u'%s %s\n' % (name, trans))
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
usage = '''
|
||||
Usage: python %(exec)s (audio_directory|in.scp) decode_directory
|
||||
|
||||
Where directory contains files "*.scp" and
|
||||
audio files "*.wav" and their transcriptions "*.wav.trn".
|
||||
The "*.scp" files contains of list wav names and their path.
|
||||
|
||||
The %(exec)s looks for "*.scp" files builds a reference from "*.wav.trn"
|
||||
'''
|
||||
usage_args = {'exec': sys.argv[0]}
|
||||
|
||||
if len(sys.argv) != 3:
|
||||
print >> sys.stderr, "Wrong number of arguments"
|
||||
print >> sys.stderr, usage % {'exec': sys.argv[0]}
|
||||
sys.exit(1)
|
||||
|
||||
if sys.argv[1].endswith('scp'):
|
||||
scps = [sys.argv[1]]
|
||||
else:
|
||||
scps = glob.glob(os.path.join(sys.argv[1], '*.scp'))
|
||||
target_dir = sys.argv[2]
|
||||
if not len(scps):
|
||||
print >> sys.stderr, "No '*.scp' files found"
|
||||
print >> sys.stderr, usage % {'exec': sys.argv[0]}
|
||||
sys.exit(1)
|
||||
if not os.path.isdir(target_dir):
|
||||
print >> sys.stderr, "No '*.scp' files found"
|
||||
print >> sys.stderr, usage % {'exec': sys.argv[0]}
|
||||
sys.exit(1)
|
||||
|
||||
refers = [os.path.join(target_dir, os.path.basename(scp) + '.tra') for scp in scps]
|
||||
for scp, refer in zip(scps, refers):
|
||||
build_reference(scp, refer)
|
||||
|
|
@ -0,0 +1 @@
|
|||
*
|
||||
|
|
@ -0,0 +1 @@
|
|||
*
|
||||
|
|
@ -0,0 +1,8 @@
|
|||
#!/bin/bash
|
||||
|
||||
# source the settings
|
||||
. path.sh
|
||||
|
||||
for n in `cut -d' ' -f1 $wav_scp` ; do
|
||||
utils/show_lattice.sh --mode save --format svg $n $lattice $wst
|
||||
done
|
||||
|
|
@ -0,0 +1,135 @@
|
|||
#!/usr/bin/env python
|
||||
# encoding: utf-8
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
from __future__ import unicode_literals
|
||||
|
||||
import pyaudio
|
||||
from kaldi.decoders import PyOnlineLatgenRecogniser
|
||||
from kaldi.utils import wst2dict, lattice_to_nbest
|
||||
import sys
|
||||
import time
|
||||
import select
|
||||
import tty
|
||||
import termios
|
||||
import wave
|
||||
|
||||
CHANNELS, RATE, FORMAT = 1, 16000, pyaudio.paInt16
|
||||
|
||||
|
||||
class LiveDemo:
|
||||
|
||||
def __init__(self, audio_batch_size, wst, dec_args):
|
||||
self.batch_size = audio_batch_size
|
||||
self.wst = wst
|
||||
self.args = dec_args
|
||||
self.d = PyOnlineLatgenRecogniser()
|
||||
self.pin, self.stream = None, None
|
||||
self.frames = []
|
||||
self.utt_frames, self.new_frames = 0, 0
|
||||
self.utt_end, self.dialog_end = False, False
|
||||
|
||||
def setup(self):
|
||||
self.d.reset()
|
||||
self.d.setup(argv)
|
||||
self.pin = pyaudio.PyAudio()
|
||||
self.stream = self.pin.open(format=FORMAT, channels=CHANNELS,
|
||||
rate=RATE, input=True, frames_per_buffer=self.batch_size,
|
||||
stream_callback=self.get_audio_callback())
|
||||
self.utt_frames, self.new_frames = 0, 0
|
||||
self.utt_end, self.dialog_end = False, False
|
||||
self.frames = []
|
||||
|
||||
def tear_down(self):
|
||||
if self.stream is not None:
|
||||
self.stream.stop_stream()
|
||||
self.stream.close()
|
||||
if self.pin is not None:
|
||||
self.pin.terminate()
|
||||
p, stream = None, None
|
||||
self.frames = []
|
||||
|
||||
def get_audio_callback(self):
|
||||
def frame_in(in_data, frame_count, time_info, status):
|
||||
self.d.frame_in(in_data)
|
||||
self.frames.append(in_data)
|
||||
return in_data, pyaudio.paContinue
|
||||
return frame_in
|
||||
|
||||
def _user_control(self):
|
||||
'''Simply stupid sollution how to control state of recogniser.'''
|
||||
|
||||
self.utt_end, self.dialog_end = False, False
|
||||
old_settings = termios.tcgetattr(sys.stdin)
|
||||
try:
|
||||
tty.setcbreak(sys.stdin.fileno())
|
||||
# if is data on input
|
||||
while (select.select([sys.stdin], [], [], 1) == ([sys.stdin], [], [])):
|
||||
c = sys.stdin.read(1)
|
||||
if c == 'u':
|
||||
print('\nMarked end of utterance\n')
|
||||
self.utt_end = True
|
||||
elif c == 'c':
|
||||
self.dialog_end = True
|
||||
print('\nMarked end of dialogue\n')
|
||||
finally:
|
||||
termios.tcsetattr(sys.stdin, termios.TCSADRAIN, old_settings)
|
||||
print("""Chunks: %d ; Utterance %d ; end %d : press 'u'\nFor terminating press 'c'\n\n""" % (len(self.frames), self.utt_frames, self.utt_end))
|
||||
|
||||
def run(self):
|
||||
while True:
|
||||
time.sleep(0.1)
|
||||
self._user_control()
|
||||
new_frames = self.d.decode(max_frames=10)
|
||||
while new_frames > 0:
|
||||
self.utt_frames += new_frames
|
||||
new_frames = self.d.decode(max_frames=10)
|
||||
if self.utt_end or self.dialog_end:
|
||||
start = time.time()
|
||||
self.d.prune_final()
|
||||
prob, lat = self.d.get_lattice()
|
||||
# lat.write('live-demo-recorded.fst')
|
||||
nbest = lattice_to_nbest(lat, n=10)
|
||||
if nbest:
|
||||
best_prob, best_path = nbest[0]
|
||||
decoded = ' '.join([wst[w] for w in best_path])
|
||||
else:
|
||||
decoded = 'Empty hypothesis'
|
||||
print("%s secs, frames: %d, prob: %f, %s " % (
|
||||
str(time.time() - start), self.utt_frames, prob, decoded))
|
||||
self.utt_frames = 0
|
||||
self.d.reset(keep_buffer_data=False)
|
||||
if self.dialog_end:
|
||||
self.save_wav()
|
||||
break
|
||||
|
||||
def save_wav(self):
|
||||
wf = wave.open('live-demo-record.wav', 'wb')
|
||||
wf.setnchannels(CHANNELS)
|
||||
wf.setframerate(RATE)
|
||||
wf.setsampwidth(self.pin.get_sample_size(FORMAT))
|
||||
wf.writeframes(b''.join(self.frames))
|
||||
wf.close()
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
audio_batch_size, wst_path = int(sys.argv[1]), sys.argv[2]
|
||||
argv = sys.argv[3:]
|
||||
print >> sys.stderr, 'Python args: %s' % str(sys.argv)
|
||||
|
||||
wst = wst2dict(wst_path)
|
||||
demo = LiveDemo(audio_batch_size, wst, argv)
|
||||
demo.setup()
|
||||
demo.run()
|
||||
|
|
@ -0,0 +1 @@
|
|||
*
|
||||
|
|
@ -0,0 +1,41 @@
|
|||
#!/bin/bash
|
||||
|
||||
# data location
|
||||
PWD=`pwd`
|
||||
exp_dir=$PWD
|
||||
data_dir=$PWD/data/vystadial-sample-cs/test
|
||||
decode_dir=$exp_dir/decode
|
||||
|
||||
# IO parameters
|
||||
wav_scp=$data_dir/input_best.scp
|
||||
# wav_scp=$data_dir/input.scp
|
||||
|
||||
gmm_latgen_faster_tra=$decode_dir/gmm-latgen-faster.tra
|
||||
gmm_latgen_faster_tra_txt=${gmm_latgen_faster_tra}.txt
|
||||
|
||||
pykaldi_latgen_tra=$decode_dir/pykaldi-latgen.tra
|
||||
pykaldi_latgen_tra_txt=${pykaldi_faster_tra}.txt
|
||||
lattice=$decode_dir/lat.gz
|
||||
|
||||
# Czech language models
|
||||
LANG=cs
|
||||
HCLG=models/HCLG_tri2b_bmmi.fst
|
||||
AM=models/tri2b_bmmi.mdl
|
||||
MAT=models/tri2b_bmmi.mat # matrix trained in tri2b models
|
||||
WST=models/words.txt
|
||||
MFCC=models/mfcc.conf
|
||||
SILENCE=models/silence.csl
|
||||
|
||||
kaldisrc=`pwd`/../../../src
|
||||
openfst=`pwd`/../../../tools/openfst/
|
||||
|
||||
export PATH=$kaldisrc/bin:$kaldisrc/fgmmbin:$kaldisrc/gmmbin:$kaldisrc/nnetbin:$kaldisrc/sgmm2bin:$kaldisrc/tiedbin:$kaldisrc/featbin:$kaldisrc/fstbin:$kaldisrc/latbin:$kaldisrc/onlinebin:$kaldisrc/sgmmbin:$kaldisrc/onl-rec:$openfst/bin:"$PATH"
|
||||
export LD_LIBRARY_PATH=$kaldisrc/onl-rec:$kaldisrc/pykaldi/kaldi:$openfst/lib:$openfst/lib/fst:$LD_LIBRARY_PATH
|
||||
export PYTHONPATH=$kaldisrc/pykaldi:$kaldisrc/pykaldi/pyfst:$PYTHONPATH
|
||||
|
||||
beam=16.0
|
||||
latbeam=10.0
|
||||
max_active=14000
|
||||
|
||||
# Size of chunks are queued in "online" interface
|
||||
batch_size=4560
|
||||
|
|
@ -0,0 +1,100 @@
|
|||
#!/usr/bin/env python
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
from __future__ import unicode_literals
|
||||
|
||||
from kaldi.utils import load_wav, wst2dict, lattice_to_nbest
|
||||
from kaldi.decoders import PyOnlineLatgenRecogniser
|
||||
import sys
|
||||
import fst
|
||||
import time
|
||||
|
||||
# DEBUG = True
|
||||
DEBUG = False
|
||||
|
||||
|
||||
def write_decoded(f, wav_name, word_ids, wst):
|
||||
assert(len(word_ids) > 0)
|
||||
best_weight, best_path = word_ids[0]
|
||||
if wst is not None:
|
||||
decoded = [wst[w] for w in best_path]
|
||||
else:
|
||||
decoded = [unicode(w) for w in best_path]
|
||||
line = u' '.join([wav_name] + decoded + ['\n'])
|
||||
if DEBUG:
|
||||
print '%s best path %s' % (wav_name, decoded.encode('UTF-8'))
|
||||
for i, s in enumerate(word_ids):
|
||||
if i > 0:
|
||||
break
|
||||
print 'best path %d: %s' % (i, str(s))
|
||||
f.write(line.encode('UTF-8'))
|
||||
|
||||
|
||||
# @profile
|
||||
def decode(d, pcm):
|
||||
frame_len = (2 * audio_batch_size) # 16-bit audio so 1 sample = 2 chars
|
||||
i, decoded_frames, max_end = 0, 0, len(pcm)
|
||||
start = time.time()
|
||||
while i * frame_len < len(pcm):
|
||||
i, begin, end = i + 1, i * frame_len, min(max_end, (i + 1) * frame_len)
|
||||
audio_chunk = pcm[begin:end]
|
||||
d.frame_in(audio_chunk)
|
||||
dec_t = d.decode(max_frames=10)
|
||||
while dec_t > 0:
|
||||
decoded_frames += dec_t
|
||||
dec_t = d.decode(max_frames=10)
|
||||
print "forward decode: %s secs" % str(time.time() - start)
|
||||
start = time.time()
|
||||
d.prune_final()
|
||||
lik, lat = d.get_lattice()
|
||||
print "backward decode: %s secs" % str(time.time() - start)
|
||||
d.reset(keep_buffer_data=False)
|
||||
return (lat, lik, decoded_frames)
|
||||
|
||||
|
||||
def decode_wrap(argv, audio_batch_size, wav_paths,
|
||||
file_output, wst_path=None):
|
||||
wst = wst2dict(wst_path)
|
||||
d = PyOnlineLatgenRecogniser()
|
||||
d.setup(argv)
|
||||
for wav_name, wav_path in wav_paths:
|
||||
sw, sr = 2, 16000 # 16-bit audio so 1 sample_width = 2 chars
|
||||
pcm = load_wav(wav_path, def_sample_width=sw, def_sample_rate=sr)
|
||||
print '%s has %f sec' % (wav_name, (float(len(pcm)) / sw) / sr)
|
||||
lat, lik, decoded_frames = decode(d, pcm)
|
||||
lat.isyms = lat.osyms = fst.read_symbols_text(wst_path)
|
||||
if DEBUG:
|
||||
with open('pykaldi_%s.svg' % wav_name, 'w') as f:
|
||||
f.write(lat._repr_svg_())
|
||||
lat.write('%s_pykaldi.fst' % wav_name)
|
||||
|
||||
print "Log-likelihood per frame for utterance %s is %f over %d frames" % (
|
||||
wav_name, (lik / decoded_frames), decoded_frames)
|
||||
word_ids = lattice_to_nbest(lat, n=10)
|
||||
write_decoded(file_output, wav_name, word_ids, wst)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
audio_scp, audio_batch_size = sys.argv[1], int(sys.argv[2])
|
||||
dec_hypo, wst_path = sys.argv[3], sys.argv[4]
|
||||
argv = sys.argv[5:]
|
||||
print >> sys.stderr, 'Python args: %s' % str(sys.argv)
|
||||
|
||||
# open audio_scp, decode and write to dec_hypo file
|
||||
with open(audio_scp, 'rb') as r:
|
||||
with open(dec_hypo, 'wb') as w:
|
||||
lines = r.readlines()
|
||||
scp = [tuple(line.strip().split(' ', 1)) for line in lines]
|
||||
decode_wrap(argv, audio_batch_size, scp, w, wst_path)
|
||||
|
|
@ -0,0 +1,55 @@
|
|||
# $!/bin/bash
|
||||
|
||||
# source the settings
|
||||
. path.sh
|
||||
|
||||
. utils/parse_options.sh || exit 1
|
||||
|
||||
# temporary files
|
||||
mfccdir=$decode_dir/mfcc
|
||||
feat_scp=$mfccdir/raw_mfcc.scp
|
||||
|
||||
mkdir -p $mfccdir
|
||||
|
||||
compute-mfcc-feats --verbose=0 --config=$MFCC scp:$wav_scp \
|
||||
ark,scp:$mfccdir/raw_mfcc.ark,$feat_scp || exit 1;
|
||||
|
||||
# # For debugging
|
||||
# # cgdb -q -x .gdbinit --args \
|
||||
# compute-mfcc-feats --verbose=0 --config=$MFCC scp:$wav_scp \
|
||||
# ark,t,scp:$mfccdir/raw_mfcc.ark.txt,${feat_scp}.txt || exit 1;
|
||||
# # For debugging
|
||||
# add-deltas "scp,s,cs:$feat_scp" "ark,t:$mfccdir/dd_mfcc.ark.txt"
|
||||
|
||||
# check the splice-feats use the default left=right=4 context
|
||||
if [ -z $MAT ] ; then
|
||||
# no LDA matrix -> use delta-delta
|
||||
feats="ark,s,cs:copy-feats scp:$feat_scp ark:- | add-deltas ark:- ark:- |"
|
||||
else
|
||||
# LDA matrix specified -> using it
|
||||
feats="ark,s,cs:copy-feats scp:$feat_scp ark:- | splice-feats ark:- ark:- | transform-feats $MAT ark:- ark:- |"
|
||||
fi
|
||||
|
||||
gmm-latgen-faster --verbose=0 --max-mem=500000000 \
|
||||
--beam=$beam --lattice-beam=$latbeam --max-active=$max_active \
|
||||
--allow-partial=true --word-symbol-table=$WST \
|
||||
$AM $HCLG "$feats" \
|
||||
"ark:|gzip -c > $lattice"
|
||||
|
||||
lattice-best-path --verbose=0 --lm-scale=15 --word-symbol-table=$WST \
|
||||
"ark:gunzip -c $lattice|" ark,t:$gmm_latgen_faster_tra || exit 1;
|
||||
|
||||
cat $gmm_latgen_faster_tra | utils/int2sym.pl -f 2- $WST \
|
||||
> $gmm_latgen_faster_tra_txt || exit 1
|
||||
|
||||
# reference is named based on wav_scp
|
||||
./build_reference.py $wav_scp $decode_dir
|
||||
reference=$decode_dir/`basename $wav_scp`.tra
|
||||
|
||||
echo; echo "Reference"; echo
|
||||
cat $reference
|
||||
echo; echo "Decoded"; echo
|
||||
cat $gmm_latgen_faster_tra_txt
|
||||
|
||||
compute-wer --text --mode=present ark:$reference ark,p:$gmm_latgen_faster_tra_txt
|
||||
|
||||
|
|
@ -0,0 +1,16 @@
|
|||
#!/bin/bash
|
||||
|
||||
# source the settings
|
||||
. path.sh
|
||||
|
||||
batch_size=4560
|
||||
beam=12.0
|
||||
latbeam=6.0
|
||||
max_active=2000
|
||||
|
||||
# cgdb -q -x .gdbinit_faster --args python \
|
||||
python \
|
||||
live-demo.py $batch_size $WST \
|
||||
--verbose=0 --lat-lm-scale=15 --config=$MFCC \
|
||||
--beam=$beam --lattice-beam=$latbeam --max-active=$max_active \
|
||||
$AM $HCLG `cat $SILENCE` $MAT
|
||||
|
|
@ -0,0 +1,18 @@
|
|||
#!/bin/bash
|
||||
|
||||
# source the settings
|
||||
. path.sh
|
||||
|
||||
. utils/parse_options.sh || exit 1
|
||||
|
||||
|
||||
# cgdb -q -x .gdbinit_latgen --args \
|
||||
|
||||
wav_name=./data/vystadial-sample-cs/test/vad-2013-06-08-22-50-01.897179.wav
|
||||
onl-rec-latgen-recogniser-test $wav_name \
|
||||
--verbose=0 --max-mem=500000000 --lat-lm-scale=15 --config=$MFCC \
|
||||
--beam=$beam --lattice-beam=$latbeam --max-active=$max_active \
|
||||
$AM $HCLG `cat $SILENCE` $MAT
|
||||
|
||||
echo; echo "Converting the lattice to svg picture ${wav_name}.svg" ; echo
|
||||
fstdraw --portrait=true --osymbols=$WST ${wav_name}.fst | dot -Tsvg > ${wav_name}.svg
|
||||
|
|
@ -0,0 +1,37 @@
|
|||
#!/bin/bash
|
||||
|
||||
# source the settings
|
||||
. path.sh
|
||||
|
||||
. utils/parse_options.sh || exit 1
|
||||
|
||||
logname=b${beam}_lb${latbeam}_ma${max_active}_bs${batch_size}
|
||||
|
||||
# Below, there are various commands for debugging, profiling.
|
||||
# Uncomment convenient prefix for you and put it just before the arguments.
|
||||
#
|
||||
# cgdb -q -x .gdbinit_latgen --args python \
|
||||
# valgrind --tool=callgrind -v --dump-instr=yes --trace-jump=yes --callgrind-out-file=callgrind_${logname}.log python \
|
||||
# kernprof.py -o kernprof_${logname}.log -l -v \
|
||||
python \
|
||||
pykaldi-online-latgen-recogniser.py $wav_scp $batch_size $pykaldi_latgen_tra $WST \
|
||||
--verbose=0 --max-mem=500000000 --lat-lm-scale=15 --config=$MFCC \
|
||||
--beam=$beam --lattice-beam=$latbeam --max-active=$max_active \
|
||||
$AM $HCLG `cat $SILENCE` $MAT
|
||||
|
||||
# If using callgrind display the results by running kcachegrind
|
||||
# kcachegrind callgrind_${logname}.log
|
||||
# If using kernprof.py @profile decorators
|
||||
# to functions which should be profiled.
|
||||
|
||||
# reference is named based on wav_scp
|
||||
./build_reference.py $wav_scp $decode_dir
|
||||
reference=$decode_dir/`basename $wav_scp`.tra
|
||||
|
||||
echo; echo "Reference"; echo
|
||||
cat $reference
|
||||
echo; echo "Decoded"; echo
|
||||
cat $pykaldi_latgen_tra
|
||||
echo
|
||||
|
||||
compute-wer --text --mode=present ark:$reference ark,p:$pykaldi_latgen_tra
|
||||
|
|
@ -0,0 +1 @@
|
|||
../../wsj/s5/utils
|
||||
|
|
@ -0,0 +1,22 @@
|
|||
# "queue.pl" uses qsub. The options to it are
|
||||
# options to qsub. If you have GridEngine installed,
|
||||
# change this to a queue you have access to.
|
||||
# Otherwise, use "run.pl", which will run jobs locally
|
||||
# (make sure your --num-jobs options are no more than
|
||||
# the number of cpus on your machine.
|
||||
|
||||
#export train_cmd="queue.pl -q all.q@a*.clsp.jhu.edu"
|
||||
#export decode_cmd="queue.pl -q all.q@a*.clsp.jhu.edu"
|
||||
# export train_cmd="queue.pl -l mf=5g"
|
||||
# export decode_cmd="queue.pl -l mf=5g"
|
||||
export train_cmd="queue.pl -l arch=*64*"
|
||||
export decode_cmd="queue.pl -l arch=*64*"
|
||||
|
||||
# The number of parallel jobs to be started for some parts of the recipe
|
||||
# Make sure you have enough resources(CPUs and RAM) to accomodate this number of jobs
|
||||
njobs=20
|
||||
|
||||
# If you have no GridEngine you can do:
|
||||
#export train_cmd=run.pl
|
||||
#export decode_cmd=run.pl
|
||||
#njobs=2
|
||||
|
|
@ -0,0 +1,3 @@
|
|||
beam=12.0
|
||||
latbeam=6.0
|
||||
max_active=14000
|
||||
|
|
@ -0,0 +1,7 @@
|
|||
# --use-energy=false # non default option. False -> Use C0 instead energy
|
||||
# NUMCEPS in HTK is without C0: there 12 here 13 - default
|
||||
--low-freq=125
|
||||
--high-freq=3800
|
||||
# --htk-compat
|
||||
# --remove-dc-offset # equivalent ZMEANSOURCE in HTK
|
||||
# --subtract-mean # not recommended to do it this way
|
||||
|
|
@ -0,0 +1,42 @@
|
|||
#!/bin/bash
|
||||
|
||||
# EVERY_N utterance is used for training
|
||||
# EVERY_N=3 -> we use one third of training data
|
||||
export EVERY_N=1
|
||||
export TEST_SETS="dev test"
|
||||
|
||||
# Directories set up
|
||||
export DATA_ROOT=`pwd`/data # expects subdirectories train + $TEST_SETS
|
||||
export WORK=`pwd`/lang_prep
|
||||
export EXP=`pwd`/exp
|
||||
export TGT_MODELS=trained_models
|
||||
|
||||
# Specify paths to arpa models. Paths may not contain spaces.
|
||||
# Specify build0 or build1 or build2, .. for building (zero|uni|bi)-gram LM.
|
||||
# Note: The LM file name should not contain underscore "_"!
|
||||
# Otherwise the results will be reported without the LM with underscore."
|
||||
export LMs="build0 build2"
|
||||
|
||||
# Use path to prebuilt dictionary or 'build' command in order to build dictionary
|
||||
# export DICTIONARY="../../resources/lm/caminfo/dict"
|
||||
export DICTIONARY="build"
|
||||
|
||||
|
||||
# Borders for estimating LM model weight.
|
||||
# LMW is tuned on development set and applied on test set.
|
||||
export min_lmw=4
|
||||
export max_lmw=15
|
||||
|
||||
# Number of states for phonem training
|
||||
export pdf=1200
|
||||
|
||||
# Maximum number of Gaussians used for training
|
||||
export gauss=19200
|
||||
|
||||
export train_mmi_boost=0.05
|
||||
|
||||
export mmi_beam=16.0
|
||||
export mmi_lat_beam=10.0
|
||||
|
||||
# --fake -> NO CMVN; empty -> CMVN (pykaldi decoders can not handle CMVN -> fake)
|
||||
export fake="--fake"
|
||||
|
|
@ -0,0 +1,41 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
expdir=$1; shift
|
||||
tgtdir=$2; shift
|
||||
|
||||
tgtdir="$tgtdir/$name"
|
||||
date="`date +%F_%T.%N`"
|
||||
|
||||
if [[ -d $tgtdir || -f $tgtdir ]] ; then
|
||||
tgtdir="$tgtdir/backup_$date"
|
||||
fi
|
||||
|
||||
|
||||
# This is EXAMPLE SCRIPT you are ENCOURAGED TO CHANGE IT!
|
||||
|
||||
mkdir -p "$tgtdir"
|
||||
cp -rf $expdir "$tgtdir"
|
||||
|
||||
# Collect the results
|
||||
|
||||
local/results.py $EXP > "$tgtdir"/results.log
|
||||
echo "Date: $date" >> "$tgtdir"/results.log
|
||||
size=`du -hs "$tgtdir"`
|
||||
echo "Size of backup: $size" >> "$tgtdir"/results.log
|
||||
|
||||
echo; echo "DATA successfully copied to $tgtdir"; echo
|
||||
|
|
@ -0,0 +1,68 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright 2012 Vassil Panayotov
|
||||
# 2013 Ondrej Platek
|
||||
# Apache 2.0
|
||||
|
||||
echo "===test_sets Formating data ..."
|
||||
langdir=$1; shift
|
||||
LMs=$1; shift
|
||||
lmdir=$1; shift
|
||||
lexicon=$1; shift
|
||||
|
||||
# Next, for each type of language model, create the corresponding FST
|
||||
# and the corresponding lang_test_* directory.
|
||||
for lm in $LMs ; do
|
||||
tgt=${langdir}_`basename "$lm"`
|
||||
lmp=$lmdir/`basename $lm`
|
||||
|
||||
tmpdir=$tgt/tmp
|
||||
mkdir -p $tgt
|
||||
mkdir -p $tmpdir
|
||||
|
||||
echo "--- Preparing the grammar transducer (G.fst) from $lmp in $tgt ..."
|
||||
|
||||
for f in phones.txt words.txt phones.txt L.fst L_disambig.fst phones ; do
|
||||
ln -s $langdir/$f $tgt/$f 2> /dev/null
|
||||
done
|
||||
|
||||
cat $lmp | utils/find_arpa_oovs.pl $tgt/words.txt > $tmpdir/oovs.txt
|
||||
|
||||
# grep -v '<s> <s>' because the LM seems to have some strange and useless
|
||||
# stuff in it with multiple <s>'s in the history. Encountered some other similar
|
||||
# things in a LM from Geoff. Removing all "illegal" combinations of <s> and </s>,
|
||||
# which are supposed to occur only at being/end of utt. These can cause
|
||||
# determinization failures of CLG [ends up being epsilon cycles].
|
||||
|
||||
cat $lmp | \
|
||||
grep -v '<s> <s>\|</s> <s>\|</s> </s>' | \
|
||||
arpa2fst - | fstprint | \
|
||||
utils/remove_oovs.pl $tmpdir/oovs.txt | \
|
||||
utils/eps2disambig.pl | utils/s2eps.pl | fstcompile --isymbols=$tgt/words.txt \
|
||||
--osymbols=$tgt/words.txt --keep_isymbols=false --keep_osymbols=false | \
|
||||
fstrmepsilon > $tgt/G.fst
|
||||
fstisstochastic $tgt/G.fst
|
||||
# The output is like:
|
||||
# 9.14233e-05 -0.259833
|
||||
# we do expect the first of these 2 numbers to be close to zero (the second is
|
||||
# nonzero because the backoff weights make the states sum to >1).
|
||||
# Because of the <s> fiasco for these particular LMs, the first number is not
|
||||
# as close to zero as it could be.
|
||||
|
||||
# Everything below is only for diagnostic.
|
||||
# Checking that G has no cycles with empty words on them (e.g. <s>, </s>);
|
||||
# this might cause determinization failure of CLG.
|
||||
# #0 is treated as an empty word.
|
||||
mkdir -p $tmpdir/g
|
||||
awk '{if(NF==1){ printf("0 0 %s %s\n", $1,$1); }} END{print "0 0 #0 #0"; print "0";}' \
|
||||
< "$lexicon" >$tmpdir/g/select_empty.fst.txt
|
||||
fstcompile --isymbols=$tgt/words.txt --osymbols=$tgt/words.txt \
|
||||
$tmpdir/g/select_empty.fst.txt | \
|
||||
fstarcsort --sort_type=olabel | fstcompose - $tgt/G.fst > $tmpdir/g/empty_words.fst
|
||||
fstinfo $tmpdir/g/empty_words.fst | grep cyclic | grep -w 'y' &&
|
||||
echo "Language model has cycles with empty words" && exit 1
|
||||
|
||||
# rm -rf $tmpdir # TODO debugging
|
||||
echo "*** Succeeded in creating G.fst for $tgt"
|
||||
|
||||
done # for lm in $LMs ; do
|
||||
|
|
@ -0,0 +1,89 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
locdata=$1; shift
|
||||
train_text=$1; shift
|
||||
test_text=$1; shift
|
||||
local_lm=$1; shift
|
||||
lms=$1; shift
|
||||
|
||||
|
||||
mkdir -p $local_lm
|
||||
|
||||
echo "=== Preparing the LM ..."
|
||||
|
||||
function build_0gram {
|
||||
echo "=== Building zerogram $lm from ${transcr}. ..."
|
||||
transcr=$1; lm=$2
|
||||
cut -d' ' -f2- $transcr | tr ' ' '\n' | sort -u > $lm
|
||||
echo "<s>" >> $lm
|
||||
echo "</s>" >> $lm
|
||||
python -c """
|
||||
import math
|
||||
with open('$lm', 'r+') as f:
|
||||
lines = f.readlines()
|
||||
p = math.log10(1/float(len(lines)));
|
||||
lines = ['%f\\t%s'%(p,l) for l in lines]
|
||||
f.seek(0); f.write('\\n\\\\data\\\\\\nngram 1= %d\\n\\n\\\\1-grams:\\n' % len(lines))
|
||||
f.write(''.join(lines) + '\\\\end\\\\')
|
||||
"""
|
||||
}
|
||||
|
||||
for lm in $lms ; do
|
||||
lm_base=`basename $lm`
|
||||
if [ ${lm_base%[0-6]} != 'build' ] ; then
|
||||
cp $lm $local_lm
|
||||
else
|
||||
# We will build the LM 'build[0-9].arpa
|
||||
lm_order=${lm_base#build}
|
||||
|
||||
echo "=== Building LM of order ${lm_order}..."
|
||||
if [ $lm_order -eq 0 ] ; then
|
||||
echo "Zerogram $lm_base LM is build from text: $test_text"
|
||||
cut -d' ' -f2- $test_text | sed -e 's:^:<s> :' -e 's:$: </s>:' | \
|
||||
sort -u > $locdata/lm_test.txt
|
||||
build_0gram $locdata/lm_test.txt $local_lm/${lm_base}
|
||||
else
|
||||
echo "LM $lm_base is build from text: $train_text"
|
||||
cut -d' ' -f2- $train_text | sed -e 's:^:<s> :' -e 's:$: </s>:' | \
|
||||
sort -u > $locdata/lm_train.txt
|
||||
ngram-count -text $locdata/lm_train.txt -order ${lm_order} \
|
||||
-wbdiscount -interpolate -lm $local_lm/${lm_base}
|
||||
fi
|
||||
fi
|
||||
done
|
||||
echo "*** LMs preparation finished!"
|
||||
|
||||
echo "=== Preparing the vocabulary ..."
|
||||
|
||||
if [ "$DICTIONARY" == "build" ]; then
|
||||
echo; echo "Building dictionary from train data"; echo
|
||||
cut -d' ' -f2- $train_text | tr ' ' '\n' > $locdata/vocab-full-raw.txt
|
||||
else
|
||||
echo; echo "Using predefined dictionary: ${DICTIONARY}"
|
||||
echo "Throwing away first 2 rows."; echo
|
||||
tail -n +3 $DICTIONARY | cut -f 1 > $locdata/vocab-full-raw.txt
|
||||
fi
|
||||
|
||||
echo '</s>' >> $locdata/vocab-full-raw.txt
|
||||
echo "Removing from vocabulary _NOISE_, and all '_' words from vocab-full.txt"
|
||||
cat $locdata/vocab-full-raw.txt | grep -v '_' | \
|
||||
sort -u > $locdata/vocab-full.txt
|
||||
echo "*** Vocabulary preparation finished!"
|
||||
|
||||
|
||||
echo "Removing from vocabulary _NOISE_, and all '_' words from vocab-test.txt"
|
||||
cut -d' ' -f2 $test_text | tr ' ' '\n' | grep -v '_' | sort -u > $locdata/vocab-test.txt
|
||||
|
||||
|
|
@ -0,0 +1,54 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
# The vystadial data are specific by having following marks in transcriptions
|
||||
# _INHALE_
|
||||
# _LAUGH_
|
||||
# _EHM_HMM_
|
||||
# _NOISE_
|
||||
# _EHM_HMM_
|
||||
# _SIL_
|
||||
|
||||
locdict=$1; shift
|
||||
|
||||
echo "--- Prepare nonsilence phone lists ..."
|
||||
# We suppose only nonsilence_phones in lexicon right now
|
||||
awk '{for(n=2;n<=NF;n++) { p[$n]=1; }} END{for(x in p) {print x}}' \
|
||||
$locdict/lexicon.txt | sort > $locdict/nonsilence_phones.txt
|
||||
|
||||
echo "--- Adding silence phones to lexicon ..."
|
||||
echo "_SIL_ SIL" >> $locdict/lexicon.txt
|
||||
echo "_EHM_HMM_ EHM" >> $locdict/lexicon.txt
|
||||
echo "_INHALE_ INH" >> $locdict/lexicon.txt
|
||||
echo "_LAUGH_ LAU" >> $locdict/lexicon.txt
|
||||
echo "_NOISE_ NOI" >> $locdict/lexicon.txt
|
||||
|
||||
echo "--- Sorting lexicon in place..."
|
||||
sort $locdict/lexicon.txt -o $locdict/lexicon.txt
|
||||
|
||||
echo "--- Prepare silence phone lists ..."
|
||||
echo SIL > $locdict/silence_phones.txt
|
||||
echo EHM >> $locdict/silence_phones.txt
|
||||
echo INH >> $locdict/silence_phones.txt
|
||||
echo LAU >> $locdict/silence_phones.txt
|
||||
echo NOI >> $locdict/silence_phones.txt
|
||||
|
||||
echo SIL > $locdict/optional_silence.txt
|
||||
|
||||
# Some downstream scripts expect this file exists, even if empty
|
||||
touch $locdict/extra_questions.txt
|
||||
|
||||
echo "*** Creating phone lists finished!"
|
||||
|
|
@ -0,0 +1,18 @@
|
|||
#!/bin/bash
|
||||
# example usage:
|
||||
# ./local/create_sample.sh /ha/projects/vystadial/data/asr/en/voip/ Results/vystadial-sample/ test 100
|
||||
# note that it suppose there are only *.wav and *.wav.trn and the
|
||||
# the n is the numbero of files in the directory
|
||||
|
||||
src=$1
|
||||
tgt=$2
|
||||
typ=$3 # dev/test/train
|
||||
n=$4
|
||||
|
||||
src_dir=$src/$typ
|
||||
tgt_dir=$tgt/$typ
|
||||
mkdir -p $tgt_dir
|
||||
ls $src_dir | head -n $n \
|
||||
| while read f ; do
|
||||
cp $src_dir/$f $tgt_dir
|
||||
done
|
||||
|
|
@ -0,0 +1,108 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
#
|
||||
# Makes train/test splits
|
||||
# local/voxforge_data_prep.sh --nspk_test ${nspk_test} ${SELECTED} || exit 1
|
||||
# create files: (TYPE=train|test)
|
||||
# a) ${TYPE}_trans.txt: ID transcription capitalized! No interputction
|
||||
# b) ${TYPE}_wav.scp: ID path2ID.wav
|
||||
# c) $TYPE.utt2spk: ID-recording ID-speaker
|
||||
# s) $TYPE.spk2utt
|
||||
# e) $TYPE.spk2gender all speakers are male
|
||||
# we have ID-recording = ID-speaker
|
||||
|
||||
# The vystadial data are specific by having following marks in transcriptions
|
||||
# _INHALE_
|
||||
# _LAUGH_
|
||||
# _EHM_HMM_
|
||||
# _NOISE_
|
||||
# _EHM_HMM_
|
||||
# _SIL_
|
||||
|
||||
# renice 20 $$
|
||||
|
||||
every_n=1
|
||||
|
||||
[ -f path.sh ] && . ./path.sh # source the path.
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
|
||||
if [ $# -ne 4 ] ; then
|
||||
echo "Usage: local/data_split.sh [--every-n 30] <data-directory> <local-directory> <LMs> <Test-Sets> <tgt-dir>";
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
DATA=$1; shift
|
||||
locdata=$1; shift
|
||||
LMs=$1; shift
|
||||
test_sets=$1; shift
|
||||
tgt_dir=$1; shift
|
||||
|
||||
echo "LMs $LMs test_sets $test_sets"
|
||||
|
||||
|
||||
echo "=== Starting initial Vystadial data preparation ..."
|
||||
echo "--- Making test/train data split from $DATA taking every $every_n recording..."
|
||||
|
||||
mkdir -p $locdata
|
||||
|
||||
i=0
|
||||
for s in $test_sets train ; do
|
||||
mkdir -p $locdata/$s
|
||||
ls $DATA/$s/ | sed -n /.*wav$/p |\
|
||||
while read wav ; do
|
||||
((i++)) # bash specific
|
||||
if [[ $i -ge $every_n ]] ; then
|
||||
i=0
|
||||
pwav=$DATA/$s/$wav
|
||||
trn=`cat $DATA/$s/$wav.trn`
|
||||
echo "$wav $pwav" >> $locdata/$s/wav.scp
|
||||
echo "$wav $wav" >> $locdata/$s/utt2spk
|
||||
echo "$wav $wav" >> $locdata/$s/spk2utt
|
||||
echo "$wav $trn" >> $locdata/$s/trans.txt
|
||||
# Ignoring gender -> label all recordings as male
|
||||
echo "$wav M" >> $locdata/spk2gender
|
||||
fi
|
||||
done # while read wav
|
||||
|
||||
for f in wav.scp utt2spk spk2utt trans.txt ; do
|
||||
sort "$locdata/$s/$f" -k1 -u -o "$locdata/$s/$f" # sort in place
|
||||
done # for f
|
||||
|
||||
done # for in $test_sets train
|
||||
|
||||
echo "Set 1:1 relation for spk2utt: spk in $test_sets AND train, sort in place"
|
||||
sort "$locdata/spk2gender" -k1 -o "$locdata/spk2gender"
|
||||
|
||||
echo "--- Distributing the file lists to train and ($test_sets x $LMs) directories ..."
|
||||
mkdir -p $WORK/train
|
||||
cp $locdata/train/wav.scp $WORK/train/wav.scp || exit 1;
|
||||
cp $locdata/train/trans.txt $WORK/train/text || exit 1;
|
||||
cp $locdata/train/spk2utt $WORK/train/spk2utt || exit 1;
|
||||
cp $locdata/train/utt2spk $WORK/train/utt2spk || exit 1;
|
||||
utils/filter_scp.pl $WORK/train/spk2utt $locdata/spk2gender > $WORK/train/spk2gender || exit 1;
|
||||
|
||||
for s in $test_sets ; do
|
||||
for lm in $LMs; do
|
||||
tgt_dir=$WORK/${s}_`basename ${lm}`
|
||||
mkdir -p $tgt_dir
|
||||
cp $locdata/${s}/wav.scp $tgt_dir/wav.scp || exit 1;
|
||||
cp $locdata/${s}/trans.txt $tgt_dir/text || exit 1;
|
||||
cp $locdata/${s}/spk2utt $tgt_dir/spk2utt || exit 1;
|
||||
cp $locdata/${s}/utt2spk $tgt_dir/utt2spk || exit 1;
|
||||
utils/filter_scp.pl $tgt_dir/spk2utt $locdata/spk2gender > $tgt_dir/spk2gender || exit 1;
|
||||
done
|
||||
done
|
||||
|
|
@ -0,0 +1,46 @@
|
|||
#!/bin/bash
|
||||
# Copyright Ondrej Platek Apache 2.0
|
||||
|
||||
DATA_ROOT=$1
|
||||
|
||||
url=https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11858/00-097C-0000-0023-4670-6/data_voip_cs.tgz
|
||||
name=data_voip_cs
|
||||
extract_file=106277
|
||||
|
||||
mkdir -p $DATA_ROOT
|
||||
|
||||
if [ ! -f $DATA_ROOT/${name}.tgz ] ; then
|
||||
wget $url -O $DATA_ROOT/${name}.tgz || exit 1
|
||||
echo "Data successfully downloaded"
|
||||
fi
|
||||
|
||||
if [[ ! -d $DATA_ROOT/$name && -e $DATA_ROOT/$name ]] ; then
|
||||
echo "The $DATA_ROOT/$name is not a directory and we cannot extract the data!"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ ! -d $DATA_ROOT/$name ] ; then
|
||||
mkdir $DATA_ROOT/$name
|
||||
tar xfv $DATA_ROOT/${name}.tgz -C $DATA_ROOT | \
|
||||
while read line; do
|
||||
x=$((x+1))
|
||||
echo -en "$x extracted from $extract_file files.\r"
|
||||
done
|
||||
fi
|
||||
|
||||
if [ -d $DATA_ROOT/$name ] ; then
|
||||
echo "Checking if data extracted correctly"
|
||||
num_files=`find $DATA_ROOT/$name -name '*' | wc -l`
|
||||
if [ ! $num_files -eq $extract_file ] ; then
|
||||
echo "Data extraction failed! Extracted $num_files instead of $extract_file"
|
||||
exit 1;
|
||||
fi
|
||||
echo "It seams that data are extracted correctly"
|
||||
fi
|
||||
|
||||
pushd $DATA_ROOT
|
||||
for t in test train dev ; do
|
||||
ln -s $name/$t
|
||||
done
|
||||
ln -s $name/arpa_bigram arpa-bigram
|
||||
popd
|
||||
|
|
@ -0,0 +1,50 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
tgt=$1; shift
|
||||
exp=$1; shift
|
||||
lang=$1; shift
|
||||
|
||||
mkdir -p $tgt
|
||||
|
||||
echo "--- Exporting models to $tgt ..."
|
||||
|
||||
# See local/save_check.sh which saves the settings at the beginning for details
|
||||
# cp -f $exp/alex_gitlog.log $exp/alex_gitdiff.log $exp/experiment_bash_vars.log $tgt
|
||||
cp -f $exp/experiment_bash_vars.log $tgt
|
||||
|
||||
# Store also the results
|
||||
cp -f $exp/results.log $tgt/results.log
|
||||
|
||||
|
||||
cp -f common/mfcc.conf $tgt
|
||||
|
||||
cp -f $exp/tri2a/final.mdl $tgt/tri2a.mdl
|
||||
cp -f $exp/tri2a/tree $tgt/tri2a.tree
|
||||
|
||||
cp -f $exp/tri2b/final.mdl $tgt/tri2b.mdl
|
||||
cp -f $exp/tri2b/tree $tgt/tri2b.tree
|
||||
cp -f $exp/tri2b/final.mat $tgt/tri2b.mat
|
||||
|
||||
cp -f $exp/tri2b_mmi_b*/final.mdl $tgt/tri2b_bmmi.mdl
|
||||
cp -f $exp/tri2b/tree $tgt/tri2b_bmmi.tree
|
||||
cp -f $exp/tri2b_mmi_b*/final.mat $tgt/tri2b_bmmi.mat
|
||||
|
||||
cp -f $lang/phones.txt $lang/phones/silence.csl $tgt
|
||||
|
||||
|
||||
# FIXME do I need splice_opts for something?
|
||||
|
|
@ -0,0 +1,172 @@
|
|||
#!perl -w
|
||||
|
||||
#
|
||||
# ====================================================================
|
||||
# Copyright (C) 1999-2008 Carnegie Mellon University and Alexander
|
||||
# Rudnicky. All rights reserved.
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without
|
||||
# modification, are permitted provided that the following conditions
|
||||
# are met:
|
||||
#
|
||||
# 1. Redistributions of source code must retain the above copyright
|
||||
# notice, this list of conditions and the following disclaimer.
|
||||
#
|
||||
# 2. Redistributions in binary form must reproduce the above copyright
|
||||
# notice, this list of conditions and the following disclaimer in
|
||||
# the documentation and/or other materials provided with the
|
||||
# distribution.
|
||||
#
|
||||
# This work was supported in part by funding from the Defense Advanced
|
||||
# Research Projects Agency, the Office of Naval Research and the National
|
||||
# Science Foundation of the United States of America, and by member
|
||||
# companies of the Carnegie Mellon Sphinx Speech Consortium. We acknowledge
|
||||
# the contributions of many volunteers to the expansion and improvement of
|
||||
# this dictionary.
|
||||
#
|
||||
# THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND
|
||||
# ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
|
||||
# THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
||||
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY
|
||||
# NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
||||
# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
||||
# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
||||
# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
||||
# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
||||
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
#
|
||||
# ====================================================================
|
||||
#
|
||||
|
||||
# [20050309] (air) Created.
|
||||
# strip out stress marks from a cmudict, producing a "SphinxPhones_40" dictionary
|
||||
# [20080420] (air) Changed to pass comments.
|
||||
# Fixed output collation sequence; DOS eol's
|
||||
# [20090309] (air) fixed duplicate pron and collation bugs
|
||||
# [20090331] (air) restored standard collation order (since other stuff deppends on it)
|
||||
# [20090629] (air) do not put comments into SPHINX_40 version; not all software deals
|
||||
# [20100118] (air) added $VERBOSE; this should really be a cmdline flag...
|
||||
#
|
||||
|
||||
|
||||
$VERBOSE = 0;
|
||||
|
||||
my $basecount = 0;
|
||||
my $dupl = 0;
|
||||
my $base = 0;
|
||||
my $varia = 0;
|
||||
|
||||
if ( scalar @ARGV ne 2 ) { die "usage: make_baseform <input> <output>\n"; }
|
||||
|
||||
open(IN, $ARGV[0]) || die "can't open $ARGV[0] for reading!\n";
|
||||
open(OUT,">$ARGV[1]") || die "can't open $ARGV[1] for writing!\n";
|
||||
|
||||
@header = (); # header comment lines (passed through)
|
||||
%dict = (); # words end up in here
|
||||
%histo = (); # some statistics on variants
|
||||
|
||||
get_dict(\%dict,\@header,IN); # process the entries
|
||||
|
||||
# what have we got?
|
||||
print STDERR "$basecount forms processed\n";
|
||||
print STDERR "$base baseforms, $varia variants and $dupl duplicates found.\n";
|
||||
print STDERR "variant distribution:\n";
|
||||
foreach $var ( sort keys %histo ) {
|
||||
print STDERR "$var\t$histo{$var}\n";
|
||||
}
|
||||
|
||||
# print special comments (copyright, etc.)
|
||||
# removed since it messes some things up...
|
||||
# foreach $h (@header) { print OUT "$h\n"; }
|
||||
|
||||
# print out each entry
|
||||
%dict_out = ();
|
||||
foreach $w (sort keys %dict) {
|
||||
$var=1; # variants will number starting with 2
|
||||
foreach $p ( @{$dict{$w}} ) {
|
||||
if ($var eq 1) {
|
||||
$dict_out{$w} = $p;
|
||||
$var++;
|
||||
} else {
|
||||
$dict_out{"$w($var)"} = $p;
|
||||
$var++;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
foreach $entry ( sort keys %dict_out ) {
|
||||
print OUT "$entry\t$dict_out{$entry}\n";
|
||||
}
|
||||
|
||||
close(IN);
|
||||
close(OUT);
|
||||
|
||||
#
|
||||
#
|
||||
# read in a dictionary
|
||||
sub get_dict {
|
||||
my $dict = shift; # data structure with dictionary entries
|
||||
my $header = shift;
|
||||
my $target = shift; # input file handle
|
||||
|
||||
while (<$target>) {
|
||||
s/[\r\n]+$//g; # DOS-robust chomp;
|
||||
|
||||
# process comments; blank lines ignored
|
||||
# presume that ";;; #" will be collected and emitted at the top
|
||||
if ($_ =~ /^;;; \#/) { push @$header, $_; next; } # save header info
|
||||
elsif ( $_ =~ /^;;;/ ) { next; } # ignore plain comments
|
||||
elsif ( $_ =~ /^\s*$/ ) { next; } # ignore blank lines
|
||||
|
||||
# extract the word,pron pair and prepare for processing
|
||||
($word,$pron) = /(.+?)\s+(.+?)$/;
|
||||
if (! defined $word) { print STDERR "bad entry (no head word): $_\n"; next; }
|
||||
|
||||
$basecount++;
|
||||
|
||||
if ($word =~ /\)$/) { # variant
|
||||
($root,$variant) = ($word =~ m/(.+?)\((.+?)\)/);
|
||||
} else {
|
||||
$root = $word;
|
||||
$variant = 0;
|
||||
}
|
||||
$pron = &strip_stress($pron);
|
||||
|
||||
# found a new baseform; set it up
|
||||
if ( ! defined $dict->{$root} ) {
|
||||
$dict->{$root}[0] = $pron;
|
||||
$base++;
|
||||
next;
|
||||
}
|
||||
|
||||
# old baseform; see if, after removed stress, pron is a duplicate
|
||||
foreach $var ( @{$dict->{$root}} ) {
|
||||
if ( $var eq $pron ) {
|
||||
if ($VERBOSE) {print STDERR "duplicate entry: $root ($variant) $pron\n";}
|
||||
$dupl++;
|
||||
$pron = "";
|
||||
last;
|
||||
}
|
||||
}
|
||||
|
||||
# it's a new variant on an existing baseform, keep it
|
||||
if ( $pron ne "" ) {
|
||||
push @{$dict->{$root}}, $pron;
|
||||
$varia++;
|
||||
$histo{scalar @{$dict->{$root}}}++; # track variant stats
|
||||
if ( scalar @{$dict->{$root}} > 4 ) { print STDERR "$root -- ",scalar @{$dict->{$root}},"\n"; }
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
# strip stress marks from phonetic symbols
|
||||
sub strip_stress {
|
||||
@pron = split " ", $_[0];
|
||||
my $p;
|
||||
foreach $p (@pron) { if ( $p =~ /\d$/) { $p =~ s/(\d+)$//; } }
|
||||
return ( join(" ",@pron));
|
||||
}
|
||||
|
||||
#
|
||||
|
|
@ -0,0 +1,374 @@
|
|||
#!/usr/bin/perl
|
||||
|
||||
use strict;
|
||||
use warnings;
|
||||
use utf8;
|
||||
use Encode;
|
||||
|
||||
# $ PhoneticTranscriptionCS.pl [inputFile inputFile2 ...] outputFile
|
||||
#
|
||||
# Converts Czech text in CAPITALS in utf8 to Czech phonetic alphabet in
|
||||
# utf8. All input files will be concatenated into the output file. If no
|
||||
# input files are specified, reads from STDIN.
|
||||
#
|
||||
# If you want the script to operate in another encoding, set the EV_encoding
|
||||
# environment variable to the desired encoding.
|
||||
#
|
||||
# This is a rewrite of "vyslov" shell-script by Nino Peterek and Jan Oldrich Kruza, which was using tools
|
||||
# written by Pavel Ircing. These are copy-pasted including comments into this
|
||||
# script.
|
||||
|
||||
my $enc = 'utf8';
|
||||
|
||||
my $out_fn = pop @ARGV;
|
||||
if ($out_fn) {
|
||||
close STDOUT;
|
||||
open STDOUT, '>', $out_fn or die "Couldn't open '$out_fn': $!";
|
||||
}
|
||||
|
||||
my %seen = ();
|
||||
while (<>) {
|
||||
for (decode($enc, $_)) {
|
||||
# if (/[^\w\s]/) {
|
||||
# chomp;
|
||||
# print encode($enc, $_), (' ' x 7), "sp\n";
|
||||
# next
|
||||
# }
|
||||
chomp;
|
||||
$_ = uc($_);
|
||||
|
||||
print encode($enc, $_);
|
||||
print(' ' x 7);
|
||||
exceptions();
|
||||
transcription();
|
||||
tr/[A-Z]/[a-z]/;
|
||||
prague2pilsen();
|
||||
infreq();
|
||||
|
||||
# while ($_ =~ /(.)/g) {
|
||||
# $seen{$1}++;
|
||||
# }
|
||||
|
||||
print encode($enc, $_);
|
||||
print "\n";
|
||||
}
|
||||
}
|
||||
|
||||
#print "unique chars are: ";
|
||||
#foreach (sort(keys %seen)) {
|
||||
# print encode($enc, $_);
|
||||
#}
|
||||
|
||||
sub exceptions {
|
||||
s/AA/A/g;
|
||||
|
||||
s/AKTI/AKTY/g;
|
||||
s/ANTI/ANTY/g;
|
||||
s/ARKTI/ARKTY/g;
|
||||
s/ATIK/ATYK/g;
|
||||
s/ATRAKTI/ATRAKTY/g;
|
||||
s/AUDI/AUDY/g;
|
||||
s/AUTOMATI/AUTOMATY/g;
|
||||
s/^BARRANDOV/BARANDOV/g;
|
||||
s/CAUSA/KAUZA/g;
|
||||
s/CELSIA/CELZIA/g;
|
||||
s/^CHAPLIN/ČAPLIN/g;
|
||||
s/CHIL/ČIL/g;
|
||||
s/DANIH/DANYH/g;
|
||||
s/DEALER/D ii LER/g;
|
||||
s/DIAG/DYAG/g;
|
||||
s/DIET/DYET/g;
|
||||
s/DIF/DYF/g;
|
||||
s/DIG/DYG/g;
|
||||
s/DIKT/DYKT/g;
|
||||
s/DILET/DYLET/g;
|
||||
s/DIPL/DYPL/g;
|
||||
s/DIRIG/DYRYG/g;
|
||||
s/DISK/DYSK/g;
|
||||
s/DISP/DYSP/g;
|
||||
s/DISPLAY/DYSPLEJ/g;
|
||||
s/DIST/DYST/g;
|
||||
s/DIVIDE/DYVIDE/g;
|
||||
s/DUKTI/DUKTY/g;
|
||||
s/EDIC/EDYC/g;
|
||||
s/EFEKTIV/EFEKTYV/g;
|
||||
s/ELEKTRONI/ELEKTRONY/g;
|
||||
s/ENERGETIK/ENERGETYK/g;
|
||||
s/ERROR/EROR/g;
|
||||
s/ETIK/ETYK/g;
|
||||
s/^EX([AEIOUÁÉÍÓÚŮ])/EGZ$1/g;
|
||||
s/FEMINI/FEMINY/g;
|
||||
s/FINIŠ/FINYŠ/g;
|
||||
s/FINITI/FINYTY/g;
|
||||
s/GATIV/GATYV/g;
|
||||
s/GENETI/GENETY/g;
|
||||
s/GIENI/GIENY/g;
|
||||
s/GITI/GITY/g;
|
||||
s/^GOETH/GÉT/g;
|
||||
s/IMUNI/IMUNY/g;
|
||||
s/INDIV/INDYV/g;
|
||||
s/ING/YNG/g;
|
||||
s/INICI/INYCI/g;
|
||||
s/INVESTI/INVESTY/g;
|
||||
s/KANDI/KANDY/g;
|
||||
s/KARATI/KARATY/g;
|
||||
s/KARDI/KARDY/g;
|
||||
s/KLAUS/KLAUZ/g;
|
||||
s/KOMODIT/KOMODYT/g;
|
||||
s/KOMUNI/KOMUNY/g;
|
||||
s/KONDI/KONDY/g;
|
||||
s/KONSOR/KONZOR/g;
|
||||
s/KREDIT/KREDYT/g;
|
||||
s/KRITI/KRITY/g;
|
||||
s/LEASING/L ii z ING/g;
|
||||
s/MANAG/MENEDŽ/g;
|
||||
s/MANIP/MANYP/g;
|
||||
s/MATI/MATY/g;
|
||||
s/MEDI/MEDY/g;
|
||||
s/MINI/MINY/g;
|
||||
s/MINUS/MÝNUS/g;
|
||||
s/MODERNI/MODERNY/g;
|
||||
s/MONIE/MONYE/g;
|
||||
s/MOTIV/MOTYV/g;
|
||||
s/^MOZART/MÓCART/g;
|
||||
s/^NE/NE!/g;
|
||||
s/^NEWTON/ŇŮTN/g;
|
||||
s/NIE/NYE/g;
|
||||
s/NII/NYY/g;
|
||||
s/NJ/Ň/g;
|
||||
s/NSTI/NSTY/g;
|
||||
s/^ODD/OD!D/g;
|
||||
s/^ODI(?=[^V])/ODY/g;
|
||||
s/^ODT/OT!T/g;
|
||||
s/OPTIM/OPTYM/g;
|
||||
s/ORGANI/ORGANY/g;
|
||||
s/^PANASONIC/PANASONYK/g;
|
||||
s/PANICK/PANYCK/g;
|
||||
s/^Patton/PETN/g;
|
||||
s/PEDIATR/PEDYATR/g;
|
||||
s/PERVITI/PERVITY/g;
|
||||
s/^PODD/POD!D/g;
|
||||
s/^PODT/POT!T/g;
|
||||
s/POLITI/POLITY/g;
|
||||
s/^POULI/PO!ULI/g;
|
||||
s/POZIT/POZYT/g;
|
||||
s/^PŘED(?=[^Ě])/PŘED!/g;
|
||||
s/PRIVATI/PRIVATY/g;
|
||||
s/PROSTITU/PROSTYTU/g;
|
||||
s/RADIK/RADYK/g;
|
||||
s/^RADIO/RADYO/g;
|
||||
s/^RÁDI(.)/RÁDY$1/g;
|
||||
s/RELATIV/RELATYV/g;
|
||||
s/RESTITU/RESTYTU/g;
|
||||
s/^ROCK/ROK/g;
|
||||
s/^ROZ/ROZ!/g;
|
||||
s/RUTIN/RUTYN/g;
|
||||
s/^SCHENGEN/ŠENGEN/g;
|
||||
s/^SEBE/SEBE!/g;
|
||||
s/SHOP/sz O P/g;
|
||||
s/^SHO/SCHO/g;
|
||||
s/SOFTWAR/SOFTVER/g;
|
||||
s/SORTIM/SORTYM/g;
|
||||
s/SPEKTIV/SPEKTYV/g;
|
||||
s/STATISTI/STATYSTY/g;
|
||||
s/STIK/STYK/g;
|
||||
s/STIMUL/STYMUL/g;
|
||||
s/^STROSSMAYER/ŠTROSMAJER/g;
|
||||
s/STUDI/STUDY/g;
|
||||
s/SUPERLATIV/SUPERLATYV/g;
|
||||
s/TECHNI/TECHNY/g;
|
||||
s/TELECOM/TELEKOM/g;
|
||||
s/TELEFONI/TELEFONY/g;
|
||||
s/TEMATI/TEMATY/g;
|
||||
s/^TESCO/TESKO/g;
|
||||
s/TETIK/TETYK/g;
|
||||
s/TEXTIL/TEXTYL/g;
|
||||
s/TIBET/TYBET/g;
|
||||
s/TIBOR/TYBOR/g;
|
||||
s/TICK/TYCK/g;
|
||||
s/TIRANY/TYRANY/g;
|
||||
s/TITUL/TYTUL/g;
|
||||
s/TRADI/TRADY/g;
|
||||
s/UNIVER/UNYVER/g;
|
||||
s/VENTI/VENTY/g;
|
||||
s/VERTIK/VERTYK/g;
|
||||
s/^WAGNER/WÁGNER/g;
|
||||
s/^WATT/VAT/g;
|
||||
s/^WEBBER/VEBER/g;
|
||||
s/^WEBER/VEBER/g;
|
||||
s/^WILSON/VILSON/g;
|
||||
|
||||
}
|
||||
|
||||
sub transcription {
|
||||
# namapování nechtěných znaků na model ticha
|
||||
s/^.*[0-9].*$/sil/g;
|
||||
|
||||
# náhrada víceznakových fonémů speciálním znakem, případně rozepsání znaku na více fonémů
|
||||
s/CH/#/g;
|
||||
s/W/V/g;
|
||||
s/Q/KV/g;
|
||||
s/DŽ/&/g; # v původním vyslov nefungovalo
|
||||
s/DZ/@/g;
|
||||
s/X/KS/g;
|
||||
|
||||
# ošetření Ě
|
||||
s/([BPFV])Ě/$1JE/g;
|
||||
s/DĚ/ĎE/g;
|
||||
s/TĚ/ŤE/g;
|
||||
s/NĚ/ŇE/g;
|
||||
s/MĚ/MŇE/g;
|
||||
s/Ě/E/g;
|
||||
|
||||
# změkčující i
|
||||
s/DI/ĎI/g;
|
||||
s/TI/ŤI/g;
|
||||
s/NI/ŇI/g;
|
||||
s/DÍ/ĎÍ/g;
|
||||
s/TÍ/ŤÍ/g;
|
||||
s/NÍ/ŇÍ/g;
|
||||
|
||||
# asimilace znělosti
|
||||
s/B$/P/g;
|
||||
s/B([PTŤKSŠCČ#F])/P$1/g;
|
||||
s/B([BDĎGZŽ@&H])$/P$1/g;
|
||||
s/P([BDĎGZŽ@&H])/B$1/g;
|
||||
s/D$/T/g;
|
||||
s/D([PTŤKSŠCČ#F])/T$1/g;
|
||||
s/D([BDĎGZŽ@&H])$/T$1/g;
|
||||
s/T([BDĎGZŽ@&H])/D$1/g;
|
||||
s/Ď$/Ť/g;
|
||||
s/Ď([PTŤKSŠCČ#F])/Ť$1/g;
|
||||
s/Ď([BDĎGZŽ@&H])$/Ť$1/g;
|
||||
s/Ť([BDĎGZŽ@&H])/Ď$1/g;
|
||||
s/V$/F/g;
|
||||
s/V([PTŤKSŠCČ#F])/F$1/g;
|
||||
s/V([BDĎGZŽ@&H])$/F$1/g;
|
||||
s/F([BDĎGZŽ@&H])/V$1/g;
|
||||
s/G$/K/g;
|
||||
s/G([PTŤKSŠCČ#F])/K$1/g;
|
||||
s/G([BDĎGZŽ@&H])$/K$1/g;
|
||||
s/K([BDĎGZŽ@&H])/G$1/g;
|
||||
s/Z$/S/g;
|
||||
s/Z([PTŤKSŠCČ#F])/S$1/g;
|
||||
s/Z([BDĎGZŽ@&H])$/S$1/g;
|
||||
s/S([BDĎGZŽ@&H])/Z$1/g;
|
||||
s/Ž$/Š/g;
|
||||
s/Ž([PTŤKSŠCČ#F])/Š$1/g;
|
||||
s/Ž([BDĎGZŽ@&H])$/Š$1/g;
|
||||
s/Š([BDĎGZŽ@&H])/Ž$1/g;
|
||||
s/H$/#/g;
|
||||
s/H([PTŤKSŠCČ#F])/#$1/g;
|
||||
s/H([BDĎGZŽ@&H])$/#$1/g;
|
||||
s/#([BDĎGZŽ@&H])/H$1/g;
|
||||
s/\@$/C/g;
|
||||
s/\@([PTŤKSŠCČ#F])/C$1/g;
|
||||
s/\@([BDĎGZŽ@&H])$/C$1/g;
|
||||
s/C([BDĎGZŽ@&H])/\@$1/g;
|
||||
s/&$/Č/g;
|
||||
s/&([PTŤKSŠCČ#F])/Č$1/g;
|
||||
s/&([BDĎGZŽ@&H])$/Č$1/g;
|
||||
s/Č([BDĎGZŽ@&H])/&$1/g;
|
||||
s/Ř$/>/g;
|
||||
s/Ř([PTŤKSŠCČ#F])/>$1/g;
|
||||
s/Ř([BDĎGZŽ@&H])$/>$1/g;
|
||||
s/([PTŤKSŠCČ#F])Ř/$1>/g;
|
||||
|
||||
|
||||
#zbytek
|
||||
s/NK/ng K/g;
|
||||
s/NG/ng G/g;
|
||||
s/MV/mg V/g;
|
||||
s/MF/mg F/g;
|
||||
s/NŤ/ŇŤ/g;
|
||||
s/NĎ/ŇĎ/g;
|
||||
s/NŇ/Ň/g;
|
||||
s/CC/C/g;
|
||||
s/DD/D/g;
|
||||
s/JJ/J/g;
|
||||
s/KK/K/g;
|
||||
s/LL/L/g;
|
||||
s/NN/N/g;
|
||||
s/MM/M/g;
|
||||
s/SS/S/g;
|
||||
s/TT/T/g;
|
||||
s/ZZ/Z/g;
|
||||
s/ČČ/Č/g;
|
||||
s/ŠŠ/Š/g;
|
||||
s/-//g;
|
||||
|
||||
# závěrečný přepis na HTK abecedu
|
||||
s/>/rsz /g;
|
||||
s/EU/eu /g;
|
||||
s/AU/au /g;
|
||||
s/OU/ou /g;
|
||||
s/Á/aa /g;
|
||||
s/Č/cz /g;
|
||||
s/Ď/dj /g;
|
||||
s/É/ee /g;
|
||||
s/Í/ii /g;
|
||||
s/Ň/nj /g;
|
||||
s/Ó/oo /g;
|
||||
s/Ř/rzs /g;
|
||||
s/Š/sz /g;
|
||||
s/Ť/tj /g;
|
||||
s/Ú/uu /g;
|
||||
s/Ů/uu /g;
|
||||
s/Ý/ii /g;
|
||||
s/Ž/zs /g;
|
||||
s/Y/i /g;
|
||||
s/&/dzs /g;
|
||||
s/\@/ts /g;
|
||||
s/#/ch /g;
|
||||
s/!//g;
|
||||
s/([A-Z])/$1 /g;
|
||||
|
||||
# crazy characters mapped to closest phones
|
||||
s/Ü/uu /g;
|
||||
s/Ö/o /g;
|
||||
s/Ć/ch /g;
|
||||
s/Ľ/l /g;
|
||||
s/Ś/sz /g;
|
||||
s/Ű/uu /g;
|
||||
s/Ź/zs /g;
|
||||
s/Ń/nj /g;
|
||||
s/Ę/e /g;
|
||||
s/Ě/e /g;
|
||||
s/Ĺ/l /g;
|
||||
s/Ľ/l /g;
|
||||
s/Ł/l /g;
|
||||
s/Â/a /g;
|
||||
s/Ä/a /g;
|
||||
s/Ç/c /g;
|
||||
s/Ë/e /g;
|
||||
s/Î/i /g;
|
||||
s/Ô/o /g;
|
||||
s/Ő/o /g;
|
||||
|
||||
# s/$/ sp/g;
|
||||
}
|
||||
|
||||
sub prague2pilsen {
|
||||
s/au/aw/g;
|
||||
s/ch/x/g;
|
||||
s/cz/ch/g;
|
||||
s/dzs/dzh/g;
|
||||
s/es/e s/g;
|
||||
s/eu/ew/g;
|
||||
s/ou/ow/g;
|
||||
s/rsz/rsh/g;
|
||||
s/rzs/rzh/g;
|
||||
s/sz/sh/g;
|
||||
s/ts/dz/g;
|
||||
s/zs/zh/g;
|
||||
}
|
||||
|
||||
sub infreq {
|
||||
s/dz/c/g;
|
||||
s/dzh/ch/g;
|
||||
s/ew/e u/g;
|
||||
s/mg/m/g;
|
||||
s/oo/o/g;
|
||||
}
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,21 @@
|
|||
#!/bin/bash
|
||||
|
||||
locdata=$1; shift
|
||||
locdict=$1; shift
|
||||
|
||||
|
||||
mkdir -p $locdict
|
||||
|
||||
perl local/phonetic_transcription_cs.pl $locdata/vocab-full.txt $locdict/cs_transcription.txt
|
||||
|
||||
echo "--- Searching for OOV words ..."
|
||||
gawk 'NR==FNR{words[$1]; next;} !($1 in words)' \
|
||||
$locdict/cs_transcription.txt $locdata/vocab-full.txt |\
|
||||
egrep -v '<.?s>' > $locdict/vocab-oov.txt
|
||||
|
||||
gawk 'NR==FNR{words[$1]; next;} ($1 in words)' \
|
||||
$locdata/vocab-full.txt $locdict/cs_transcription.txt |\
|
||||
egrep -v '<.?s>' > $locdict/lexicon.txt
|
||||
|
||||
wc -l $locdict/vocab-oov.txt
|
||||
wc -l $locdict/lexicon.txt
|
||||
|
|
@ -0,0 +1,187 @@
|
|||
#!/usr/bin/env python
|
||||
# -*- coding: utf-8 -*-
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
import argparse
|
||||
from numpy import mean
|
||||
import glob
|
||||
import sys
|
||||
import sqlite3
|
||||
|
||||
|
||||
def extract_stat(wer_file):
|
||||
wer, ser = None, None
|
||||
try:
|
||||
with open(wer_file, 'r') as f:
|
||||
s = f.readlines()
|
||||
wer = float(s[1].split()[1])
|
||||
ser = float(s[2].split()[1])
|
||||
|
||||
except Exception as e:
|
||||
print sys.stderr, 'Error parsing file %s' % wer_file
|
||||
print sys.stderr, str(e)
|
||||
return wer, ser
|
||||
|
||||
|
||||
def extractResults(path):
|
||||
wer_files = glob.glob('%s/*/decode_*/*wer_*' % path)
|
||||
table = []
|
||||
for wf in wer_files:
|
||||
try:
|
||||
exp, decode_dir, wer_f = wf.split('/')[-3:]
|
||||
# last split: decode_it3_dev_build0 -> (dev, build0)
|
||||
lm = decode_dir.split('_')[-1]
|
||||
dataset = decode_dir.split('_')[-2]
|
||||
lm_w = int(wer_f[4:]) # strip wer_ from wer_19
|
||||
wer, ser = extract_stat(wf)
|
||||
table.append((exp, dataset, lm, lm_w, wer, ser))
|
||||
except Exception as e:
|
||||
print >> sys.stderr, 'failed to parse %s' % wf
|
||||
print >> sys.stderr, str(e)
|
||||
return table
|
||||
|
||||
|
||||
class Table(object):
|
||||
|
||||
def __init__(self, data=[], colnames=[]):
|
||||
self.data = data
|
||||
self.colnames = colnames
|
||||
self.colSep = '\t'
|
||||
self.lineSep = '\n'
|
||||
|
||||
def data2str(self):
|
||||
strdata = []
|
||||
for r in self.data:
|
||||
strdata.append([str(c) for c in r])
|
||||
return strdata
|
||||
|
||||
def __str__(self):
|
||||
sd = self.data2str()
|
||||
colwidth = [len(c) for c in self.colnames]
|
||||
for j in range(len(colwidth)):
|
||||
for r in sd:
|
||||
colwidth[j] = max(colwidth[j], len(r[j]))
|
||||
|
||||
gaps = [m - len(c) for (m, c) in zip(colwidth, self.colnames)]
|
||||
rows = [self.colSep.join(
|
||||
[c + ' ' * gap for c, gap in zip(self.colnames, gaps)])]
|
||||
for r in sd:
|
||||
gaps = [m - len(c) for (m, c) in zip(colwidth, r)]
|
||||
rows.append(
|
||||
self.colSep.join([c + ' ' * d for c, d in zip(r, gaps)]))
|
||||
return self.lineSep.join(rows)
|
||||
|
||||
|
||||
class LatexTable(Table):
|
||||
|
||||
def __init__(self, data=[], colnames=[]):
|
||||
Table.__init__(self, data, colnames)
|
||||
nc = len(colnames)
|
||||
self.header = '\\begin{tabular}{%s}' % ('c' * nc)
|
||||
self.tail = '\\end{tabular}'
|
||||
self.colSep = ' & '
|
||||
self.lineSep = '\\\\ \n'
|
||||
|
||||
def __str__(self):
|
||||
table_s = super(LatexTable, self).__str__()
|
||||
table_s = table_s.replace('_', '\_')
|
||||
return '%s\n%s\n%s\n' % (self.header, table_s, self.tail)
|
||||
|
||||
|
||||
def Table2LatexTable(table):
|
||||
return LatexTable(table.data, table.colnames)
|
||||
|
||||
|
||||
def createSmallTable(r):
|
||||
d = []
|
||||
for k, v in r.iteritems():
|
||||
w, s, r = v
|
||||
if w == []:
|
||||
minw = None
|
||||
else:
|
||||
minw = min(w) # returns tuple if s is list of tuples
|
||||
if s == []:
|
||||
mins = None
|
||||
else:
|
||||
mins = min(s) # returns tuple if s is list of tuples
|
||||
d.append([k, mean(r), minw, mins])
|
||||
t = Table(d, ['exp', 'RT coef', 'WER', 'SER'])
|
||||
return t
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser(
|
||||
description='Parse experiment directory generated by kaldi vystadial recipe and print statistics')
|
||||
|
||||
parser.add_argument('expath', type=str, action='store',
|
||||
help='Path to experiment directory')
|
||||
parser.add_argument('-l', '--latex', default=False, action='store_true',
|
||||
help='Generate also latex format table')
|
||||
args = parser.parse_args()
|
||||
|
||||
raw_d = extractResults(args.expath)
|
||||
|
||||
conn = sqlite3.connect(':memory:')
|
||||
c = conn.cursor()
|
||||
c.execute(
|
||||
'''CREATE TABLE results (exp text, dataset text, lm text, lm_w int, wer float, ser float)''')
|
||||
c.executemany('INSERT INTO results VALUES (?, ?, ?, ?, ?, ?)', raw_d)
|
||||
|
||||
# get all results sorted
|
||||
# c.execute("SELECT * FROM results ORDER BY exp, dataset, lm, lm_w")
|
||||
# d = c.fetchall()
|
||||
# t = Table(data=d, colnames=['exp', 'set', 'lm', 'LMW', 'WER', 'SER'])
|
||||
# print '%s\n==================' % str(t)
|
||||
|
||||
# best experiment
|
||||
# c.execute("SELECT exp, dataset, lm_w, MIN(wer), ser FROM results ORDER BY exp, lm_w, dataset")
|
||||
# d = c.fetchall()
|
||||
# compare dev and test set by picking up the best experiment
|
||||
# c.execute(("SELECT exp, dataset, lm_w, MIN(wer), ser FROM results "
|
||||
# "GROUP BY exp, lm, dataset ORDER BY exp, lm, dataset"))
|
||||
# d = c.fetchall()
|
||||
# t = Table(data=d, colnames=['exp', 'set', 'lm', 'LMW', 'WER', 'SER'])
|
||||
# print '%s\n==================' % str(t)
|
||||
|
||||
# traditional usage of devset
|
||||
dev_set_query = ("SELECT r.exp, r.lm, r.lm_w FROM results AS r "
|
||||
"INNER JOIN ( SELECT dataset, exp, lm, MIN(wer) as min_wer "
|
||||
" FROM results WHERE dataset=? GROUP BY exp, lm) i "
|
||||
"ON r.exp=i.exp AND r.lm=i.lm AND r.dataset=i.dataset AND r.wer <= i.min_wer "
|
||||
)
|
||||
c.execute(dev_set_query, ('dev',))
|
||||
|
||||
min_dev = c.fetchall()
|
||||
|
||||
# remove duplicates: duplicates if equal mimimum wer in dev set
|
||||
min_dev_un = [(e, lm, lmw) for ((e, lm), lmw) in
|
||||
dict([((e, lm), lmw) for e, lm, lmw in min_dev]).items()]
|
||||
# sort according LM -> sort results according experiment & LMs
|
||||
min_dev_un.sort(key=lambda x: (x[1], x[0]))
|
||||
|
||||
# extract corresponding test results to dev set
|
||||
d = []
|
||||
for exp, lm, lm_w in min_dev_un:
|
||||
c.execute(("SELECT * FROM results WHERE "
|
||||
"dataset='test' AND exp=? AND lm=? AND lm_w=?"),
|
||||
(exp, lm, lm_w))
|
||||
x = c.fetchall()
|
||||
assert (len(x) == 1), "One row should be extracted."
|
||||
d.append(x[0])
|
||||
|
||||
t = Table(data=d, colnames=['exp', 'set', 'LM', 'LMW', 'WER', 'SER'])
|
||||
print str(t)
|
||||
if args.latex:
|
||||
print Table2LatexTable(t)
|
||||
|
|
@ -0,0 +1,49 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
EXP=$1
|
||||
|
||||
# make sure that the directories exists
|
||||
conflict=""
|
||||
for d in $@ ; do
|
||||
if [ -d $d ] || [ -f $d ] ; then
|
||||
conflict="$conflict $d"
|
||||
fi
|
||||
done
|
||||
|
||||
if [[ ! -z "$conflict" ]] ; then
|
||||
echo "Running new experiment will create following directories."
|
||||
echo "Some of them already exists!"
|
||||
echo ""
|
||||
echo "Existing directories:"
|
||||
for d in $conflict ; do
|
||||
echo " $d"
|
||||
done
|
||||
read -p "Should I delete the conflicting directories NOW y/n?"
|
||||
case $REPLY in
|
||||
[Yy]* ) echo "Deleting $conflict directories"; rm -rf $conflict;;
|
||||
* ) echo 'Keeping conflicting directories and exiting ...'; exit 1;;
|
||||
esac
|
||||
fi
|
||||
|
||||
for d in $@ ; do
|
||||
mkdir -p $d
|
||||
done
|
||||
|
||||
# Save the variables set up
|
||||
(set -o posix ; set ) > $EXP/experiment_bash_vars.log
|
||||
# git log -1 > $EXP/alex_gitlog.log
|
||||
# git diff > $EXP/alex_gitdiff.log
|
||||
|
|
@ -0,0 +1,53 @@
|
|||
#!/bin/bash
|
||||
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey)
|
||||
# 2014 Mff UK, UFAL (modification: Ondrej Platek)
|
||||
# Apache 2.0
|
||||
|
||||
[ -f ./path.sh ] && . ./path.sh
|
||||
|
||||
# begin configuration section.
|
||||
cmd=run.pl
|
||||
min_lmw=9
|
||||
max_lmw=20
|
||||
#end configuration section.
|
||||
|
||||
[ -f ./path.sh ] && . ./path.sh
|
||||
. parse_options.sh || exit 1;
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: local/score.sh [--cmd (run.pl|queue.pl...)] <data-dir> <lang-dir|graph-dir> <decode-dir>"
|
||||
echo " Options:"
|
||||
echo " --cmd (run.pl|queue.pl...) # specify how to run the sub-processes."
|
||||
echo " --min_lmw <int> # minumum LM-weight for lattice rescoring "
|
||||
echo " --max_lmw <int> # maximum LM-weight for lattice rescoring "
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
data=$1
|
||||
lang_or_graph=$2
|
||||
dir=$3
|
||||
|
||||
symtab=$lang_or_graph/words.txt
|
||||
|
||||
for f in $symtab $dir/lat.1.gz $data/text; do
|
||||
[ ! -f $f ] && echo "score.sh: no such file $f" && exit 1;
|
||||
done
|
||||
|
||||
mkdir -p $dir/scoring/log
|
||||
|
||||
cp $data/text $dir/scoring/test.txt
|
||||
|
||||
$cmd LMW=$min_lmw:$max_lmw $dir/scoring/log/best_path.LMW.log \
|
||||
lattice-best-path --lm-scale=LMW --word-symbol-table=$symtab \
|
||||
"ark:gunzip -c $dir/lat.*.gz|" ark,t:$dir/scoring/LMW.tra || exit 1;
|
||||
|
||||
$cmd LMW=$min_lmw:$max_lmw $dir/scoring/log/score.LMW.log \
|
||||
cat $dir/scoring/LMW.tra \| \
|
||||
utils/int2sym.pl -f 2- $symtab \| \
|
||||
compute-wer --text --mode=present \
|
||||
ark:$dir/scoring/test.txt ark,p:- ">&" $dir/wer_LMW || exit 1;
|
||||
|
||||
# Show results
|
||||
for f in $dir/wer_*; do echo $f; egrep '(WER)|(SER)' < $f; done
|
||||
|
||||
exit 0;
|
||||
|
|
@ -0,0 +1,19 @@
|
|||
# Needed for "correct" sorting
|
||||
export LC_ALL=C
|
||||
export KALDI_ROOT=../../..
|
||||
|
||||
# adding Kaldi binaries to path
|
||||
export PATH=$KALDI_ROOT/src/bin:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/irstlm/bin/:$KALDI_ROOT/src/fstbin/:$KALDI_ROOT/src/gmmbin/:$KALDI_ROOT/src/featbin/:$KALDI_ROOT/src/lm/:$KALDI_ROOT/src/sgmmbin/:$KALDI_ROOT/src/sgmm2bin/:$KALDI_ROOT/src/fgmmbin/:$KALDI_ROOT/src/latbin/:$PWD:$PWD/utils:$PWD/steps:$PATH
|
||||
|
||||
|
||||
|
||||
srilm_bin=$KALDI_ROOT/tools/srilm/bin/
|
||||
if [ ! -e "$srilm_bin" ] ; then
|
||||
echo "SRILM is not installed in $KALDI_ROOT/tools."
|
||||
echo "May not be able to create LMs!"
|
||||
echo "Please go to $KALDI_ROOT/tools and run ./install_srilm.sh"
|
||||
fi
|
||||
srilm_sub_bin=`find "$srilm_bin" -type d`
|
||||
for d in $srilm_sub_bin ; do
|
||||
export PATH=$d:$PATH
|
||||
done
|
||||
|
|
@ -0,0 +1,165 @@
|
|||
#!/bin/bash
|
||||
# Copyright Ondrej Platek Apache 2.0
|
||||
renice 20 $$
|
||||
|
||||
# Load training parameters
|
||||
. ./env_voip_cs.sh
|
||||
# Source optional config if exists
|
||||
[ -f env_voip_cs_CUSTOM.sh ] && . ./env_voip_cs_CUSTOM.sh
|
||||
|
||||
. ./path.sh
|
||||
|
||||
# If you have cluster of machines running GridEngine you may want to
|
||||
# change the train and decode commands in the file below
|
||||
. ./cmd.sh
|
||||
|
||||
#######################################################################
|
||||
# Preparing acoustic features, LMs and helper files #
|
||||
#######################################################################
|
||||
|
||||
echo " Copy the configuration files to $EXP directory."
|
||||
local/save_check.sh $EXP $WORK/* || exit 1;
|
||||
|
||||
local/download_cs_data.sh $DATA_ROOT || exit 1;
|
||||
|
||||
local/data_split.sh --every_n $EVERY_N $DATA_ROOT $WORK/local "$LMs" "$TEST_SETS" || exit 1
|
||||
|
||||
local/create_LMs.sh $WORK/local $WORK/local/train/trans.txt \
|
||||
$WORK/local/test/trans.txt $WORK/local/lm "$LMs" || exit 1
|
||||
|
||||
local/prepare_cs_transcription.sh $WORK/local $WORK/local/dict || exit 1
|
||||
|
||||
local/create_phone_lists.sh $WORK/local/dict || exit 1
|
||||
|
||||
utils/prepare_lang.sh $WORK/local/dict '_SIL_' $WORK/local/lang $WORK/lang || exit 1
|
||||
|
||||
local/create_G.sh $WORK/lang "$LMs" $WORK/local/lm $WORK/local/dict/lexicon.txt || exit 1
|
||||
|
||||
echo "Create MFCC features and storing them (Could be large)."
|
||||
for s in train $TEST_SETS ; do
|
||||
steps/make_mfcc.sh --mfcc-config common/mfcc.conf --cmd \
|
||||
"$train_cmd" --nj $njobs $WORK/local/$s $EXP/make_mfcc/$s $WORK/mfcc || exit 1;
|
||||
# Note --fake -> NO CMVN
|
||||
steps/compute_cmvn_stats.sh $fake $WORK/local/$s \
|
||||
$EXP/make_mfcc/$s $WORK/mfcc || exit 1;
|
||||
done
|
||||
|
||||
echo "Decoding is done for each pair (TEST_SET x LMs)"
|
||||
echo "Distribute the links to MFCC feats to all LM variations."
|
||||
cp $WORK/local/train/feats.scp $WORK/train/feats.scp
|
||||
cp $WORK/local/train/cmvn.scp $WORK/train/cmvn.scp
|
||||
for s in $TEST_SETS; do
|
||||
for lm in $LMs; do
|
||||
tgt_dir=${s}_`basename "$lm"`
|
||||
echo "cp $WORK/local/$s/feats.scp $WORK/$tgt_dir/feats.scp"
|
||||
cp $WORK/local/$s/feats.scp $WORK/$tgt_dir/feats.scp
|
||||
echo "cp $WORK/local/$s/cmvn.scp $WORK/$tgt_dir/cmvn.scp"
|
||||
cp $WORK/local/$s/cmvn.scp $WORK/$tgt_dir/cmvn.scp
|
||||
done
|
||||
done
|
||||
|
||||
#######################################################################
|
||||
# Training Acoustic Models #
|
||||
#######################################################################
|
||||
|
||||
echo "Train monophone models on full data -> may be wastefull (can be done on subset)"
|
||||
steps/train_mono.sh --nj $njobs --cmd "$train_cmd" $WORK/train $WORK/lang $EXP/mono || exit 1;
|
||||
|
||||
echo "Get alignments from monophone system."
|
||||
steps/align_si.sh --nj $njobs --cmd "$train_cmd" \
|
||||
$WORK/train $WORK/lang $EXP/mono $EXP/mono_ali || exit 1;
|
||||
|
||||
echo "Train tri1 [first triphone pass]"
|
||||
steps/train_deltas.sh --cmd "$train_cmd" \
|
||||
$pdf $gauss $WORK/train $WORK/lang $EXP/mono_ali $EXP/tri1 || exit 1;
|
||||
|
||||
# draw-tree $WORK/lang/phones.txt $EXP/tri1/tree | dot -Tsvg -Gsize=8,10.5 > graph.svg
|
||||
|
||||
echo "Align tri1"
|
||||
steps/align_si.sh --nj $njobs --cmd "$train_cmd" \
|
||||
--use-graphs true $WORK/train $WORK/lang $EXP/tri1 $EXP/tri1_ali || exit 1;
|
||||
|
||||
echo "Train tri2a [delta+delta-deltas]"
|
||||
steps/train_deltas.sh --cmd "$train_cmd" $pdf $gauss \
|
||||
$WORK/train $WORK/lang $EXP/tri1_ali $EXP/tri2a || exit 1;
|
||||
|
||||
echo "Train tri2b [LDA+MLLT]"
|
||||
steps/train_lda_mllt.sh --cmd "$train_cmd" $pdf $gauss \
|
||||
$WORK/train $WORK/lang $EXP/tri1_ali $EXP/tri2b || exit 1;
|
||||
|
||||
echo "Align all data with LDA+MLLT system (tri2b)"
|
||||
steps/align_si.sh --nj $njobs --cmd "$train_cmd" \
|
||||
--use-graphs true $WORK/train $WORK/lang $EXP/tri2b $EXP/tri2b_ali || exit 1;
|
||||
|
||||
echo "Train MMI on top of LDA+MLLT."
|
||||
steps/make_denlats.sh --nj $njobs --cmd "$train_cmd" \
|
||||
--beam $mmi_beam --lattice-beam $mmi_lat_beam \
|
||||
$WORK/train $WORK/lang $EXP/tri2b $EXP/tri2b_denlats || exit 1;
|
||||
steps/train_mmi.sh $WORK/train $WORK/lang $EXP/tri2b_ali $EXP/tri2b_denlats $EXP/tri2b_mmi || exit 1;
|
||||
|
||||
echo "Train MMI on top of LDA+MLLT with boosting. train_mmi_boost is a e.g. 0.05"
|
||||
steps/train_mmi.sh --boost ${train_mmi_boost} $WORK/train $WORK/lang \
|
||||
$EXP/tri2b_ali $EXP/tri2b_denlats $EXP/tri2b_mmi_b${train_mmi_boost} || exit 1;
|
||||
|
||||
echo "Train MPE."
|
||||
steps/train_mpe.sh $WORK/train $WORK/lang $EXP/tri2b_ali $EXP/tri2b_denlats $EXP/tri2b_mpe || exit 1;
|
||||
|
||||
#######################################################################
|
||||
# Building decoding graph #
|
||||
#######################################################################
|
||||
for lm in $LMs ; do
|
||||
lm=`basename "$lm"`
|
||||
utils/mkgraph.sh --mono $WORK/lang_${lm} $EXP/mono $EXP/mono/graph_${lm} || exit 1
|
||||
utils/mkgraph.sh $WORK/lang_${lm} $EXP/tri1 $EXP/tri1/graph_${lm} || exit 1
|
||||
utils/mkgraph.sh $WORK/lang_${lm} $EXP/tri2a $EXP/tri2a/graph_${lm} || exit 1
|
||||
utils/mkgraph.sh $WORK/lang_${lm} $EXP/tri2b $EXP/tri2b/graph_${lm} || exit 1
|
||||
done
|
||||
|
||||
|
||||
#######################################################################
|
||||
# Decoding #
|
||||
#######################################################################
|
||||
for s in $TEST_SETS ; do
|
||||
for lm in $LMs ; do
|
||||
lm=`basename "$lm"`
|
||||
tgt_dir=${s}_`basename "$lm"`
|
||||
echo "Monophone decoding"
|
||||
# Note: steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
# calls the command line once for each test,
|
||||
# and afterwards averages the WERs into (in this case $EXP/mono/decode/)
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/mono/graph_${lm} $WORK/${tgt_dir} $EXP/mono/decode_${tgt_dir}
|
||||
echo "Decode tri1"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri1/graph_${lm} $WORK/$tgt_dir $EXP/tri1/decode_${tgt_dir}
|
||||
echo "Decode tri2a"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2a/graph_${lm} $WORK/$tgt_dir $EXP/tri2a/decode_${tgt_dir}
|
||||
echo "Decode tri2b [LDA+MLLT]"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b/decode_${tgt_dir}
|
||||
# Note: change --iter option to select the best model. 4.mdl == final.mdl
|
||||
echo "Decode MMI on top of LDA+MLLT."
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --iter 4 --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b_mmi/decode_it4_${tgt_dir}
|
||||
echo "Decode MMI on top of LDA+MLLT with boosting. train_mmi_boost is a number e.g. 0.05"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --iter 4 --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b_mmi_b${train_mmi_boost}/decode_it4_${tgt_dir};
|
||||
echo "Decode MPE."
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --iter 4 --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b_mpe/decode_it4_${tgt_dir} || exit 1;
|
||||
done
|
||||
done
|
||||
|
||||
|
||||
echo "Successfully trained and evaluated all the experiments"
|
||||
local/results.py $EXP | tee $EXP/results.log
|
||||
|
||||
local/export_models.sh $TGT_MODELS $EXP $WORK/lang
|
||||
|
|
@ -0,0 +1 @@
|
|||
../../wsj/s5/steps
|
||||
|
|
@ -0,0 +1 @@
|
|||
../../wsj/s5/utils
|
||||
|
|
@ -0,0 +1,165 @@
|
|||
Summary
|
||||
-------
|
||||
The data comprise over 41 hours of speech in English.
|
||||
|
||||
The English recordings were collected from humans interacting via telephone
|
||||
calls with statistical dialogue systems, designed to provide the user
|
||||
with information on a suitable dining venue in the town.
|
||||
|
||||
The data collection process is described in detail
|
||||
in article "Free English and Czech telephone speech corpus shared under the CC-BY-SA 3.0 license"
|
||||
published for LREC 2014 (To Appear).
|
||||
|
||||
WE USE COMMON KALDI DECODERS IN THE SCRIPTS (gmm-latgen-faster through steps/decode.sh)
|
||||
However, the main purpose of providing the data and scripts
|
||||
is training acoustic models for real-time speech recognition unit
|
||||
for dialog system ALEX, which uses modified real-time Kaldi OnlineLatgenRecogniser.
|
||||
The modified Kaldi decoders are NOT required for running the scripts!
|
||||
|
||||
The modified OnlineLatgenRecogniser is actively developed at
|
||||
https://github.com/UFAL-DSG/pykaldi/tree/master/src/onl-rec
|
||||
and has Python wrapper:
|
||||
https://github.com/UFAL-DSG/pykaldi/tree/master/src/pykaldi
|
||||
Note that I am currently moving the online recogniser to:
|
||||
http://sourceforge.net/p/kaldi/code/HEAD/tree/sandbox/oplatek2/
|
||||
|
||||
Credits and license
|
||||
------------------------
|
||||
The scripts are partially based on Voxforge KALDI recipe.
|
||||
The original scripts as well as theses scripts are licensed under APACHE 2.0 license.
|
||||
The data are distributed under Attribution-{ShareAlike} 3.0 Unported ({CC} {BY}-{SA} 3.0) license.
|
||||
Czech data: https://lindat.mff.cuni.cz/repository/xmlui/handle/11858/00-097C-0000-0023-4670-6
|
||||
English data: https://lindat.mff.cuni.cz/repository/xmlui/handle/11858/00-097C-0000-0023-4671-4
|
||||
|
||||
The data collecting process and development of these training scripts
|
||||
was partly funded by the Ministry of Education, Youth and Sports
|
||||
of the Czech Republic under the grant agreement LK11221
|
||||
and core research funding of Charles University in Prague.
|
||||
For citing, please use following BibTex citation:
|
||||
|
||||
@inproceedings{korvas_2014,
|
||||
title={{Free English and Czech telephone speech corpus shared under the CC-BY-SA 3.0 license}},
|
||||
author={Korvas, Mat\v{e}j and Pl\'{a}tek, Ond\v{r}ej and Du\v{s}ek, Ond\v{r}ej and \v{Z}ilka, Luk\'{a}\v{s} and Jur\v{c}\'{i}\v{c}ek, Filip},
|
||||
booktitle={Proceedings of the Eigth International Conference on Language Resources and Evaluation (LREC 2014)},
|
||||
pages={To Appear},
|
||||
year={2014},
|
||||
}
|
||||
|
||||
|
||||
Expected results
|
||||
----------------
|
||||
The expected results were obtained simply by running
|
||||
bash train_voip_cs.sh OR bash train_voip_en.sh.
|
||||
Note that you need SRILM installed in path or at kaldi/tools/ directory!
|
||||
|
||||
build2 - bigram LM from train data, estimated by the scripts using SRILM
|
||||
build0 - zerogram LM from test data, estimated by scripts using Python code
|
||||
LMW - Language model weight, we picked the best from (min_lmw, max_lmw) based on decoding results on DEV set
|
||||
|
||||
Full English data:
|
||||
exp set LM LMW WER SER
|
||||
mono test build0 9 67.52 91.6
|
||||
tri1 test build0 10 36.75 74.7
|
||||
tri2a test build0 10 35.8 71.65
|
||||
tri2b test build0 10 32.24 68.35
|
||||
tri2b_mmi test build0 9 24.36 54.5
|
||||
tri2b_mmi_b0.05 test build0 9 23.72 53.1
|
||||
tri2b_mpe test build0 10 25.81 59.45
|
||||
mono test build2 14 31.51 64.2
|
||||
tri1 test build2 20 15.2 43.55
|
||||
tri2a test build2 16 15.61 43.4
|
||||
tri2b test build2 19 15.27 42.8
|
||||
tri2b_mmi test build2 14 10.2 30.45
|
||||
tri2b_mmi_b0.05 test build2 17 10.09 29.85
|
||||
tri2b_mpe test build2 20 15.54 42.2
|
||||
|
||||
Note that the zero-gram LMs for discriminative training
|
||||
give significant advantage, because they are estimated on test set!
|
||||
|
||||
|
||||
Details
|
||||
-------
|
||||
* Requires Kaldi installation and Linux environment. (Tested on Ubuntu 10.04, 12.04 and 12.10.)
|
||||
* The config file s5/env_voip_en.sh sets the data directory,
|
||||
mfcc directory and experiments directory.
|
||||
* Our scripts prepare the data to the expected format in s5/data.
|
||||
* Experiment files are stored to $exp directory e.g. s5/exp.
|
||||
* The local directory contains scripts for data preparation to prepare
|
||||
lang directory.
|
||||
* path.sh, cmd.sh and common/* contain configurations for the
|
||||
recipe.
|
||||
* Language model (LM) is either built from the training data using
|
||||
[SRILM](http://www.speech.sri.com/projects/srilm/) or we supply one in
|
||||
the ARPA format.
|
||||
|
||||
|
||||
Running experiments
|
||||
-------------------
|
||||
Before running the experiments, check that:
|
||||
|
||||
* you have the Kaldi toolkit compiled:
|
||||
http://sourceforge.net/projects/kaldi/.
|
||||
* you have SRILM compiled. (This is needed for building a language model
|
||||
unless you supply your own LM in the ARPA format.)
|
||||
See http://www.speech.sri.com/projects/srilm/.
|
||||
* The number of jobs njobs are set correctly in path.sh.
|
||||
* In cmd.sh, you switched to run the training on a SGE[*] grid if
|
||||
required (disabled by default).
|
||||
|
||||
Start the recipe from the s5 directory by running
|
||||
bash run.sh.
|
||||
It will create s5/mfcc, s5/data and s5/exp directories.
|
||||
If any of them exists, it will ask you if you want them to be overwritten.
|
||||
|
||||
.. [*] Sun Grid Engine
|
||||
|
||||
Extracting the results and trained models
|
||||
-----------------------------------------
|
||||
The main scripts, s5/run.sh,
|
||||
perform not only training of the acoustic models, but also decoding.
|
||||
The acoustic models are evaluated after running the training and
|
||||
reports are printed to the standard output.
|
||||
|
||||
The s5/local/results.py exp command extracts the results from the $exp directory.
|
||||
and stores the results to exp/results.log.
|
||||
|
||||
If you want to use the trained acoustic model with your language model
|
||||
outside the prepared script, you need to build the HCLG decoding graph yourself.
|
||||
See http://kaldi.sourceforge.net/graph.html for general introduction to the FST
|
||||
framework in Kaldi.
|
||||
|
||||
The simplest way to start decoding is to use the same LM which
|
||||
was used by the s5/run.sh script.
|
||||
Let's say you want to decode with
|
||||
the acoustic model stored in exp/tri2b_bmmi,
|
||||
then you need files listed below:
|
||||
|
||||
================================= =====================================================================================
|
||||
mfcc.conf Speech parametrisation (MFCC) settings. Training and decoding setup must match.
|
||||
exp/tri2b_bmmi/graph/HCLG.fst Decoding Graph. Graph part of AM plus lexicon, phone->3phone & LM representation.
|
||||
exp/tri2b_bmmi/graph/words.txt Word symbol table, a mapping between words and integers which are decoded.
|
||||
exp/tri2b_bmmi/graph/silence.csl List of phone integer ids, which represent silent phones.
|
||||
exp/tri2b_bmmi/final.mdl Trained acoustic model (AM).
|
||||
exp/tri2b_bmmi/final.mat Trained matrix of feature/space transformations (E.g. LDA and bMMI).
|
||||
================================= =====================================================================================
|
||||
|
||||
|
||||
We recommend to study steps/decode.sh Kaldi standard script
|
||||
for standalone decoding with gmm-latgen-faster Kaldi decoder.
|
||||
|
||||
In order to build your own decoding graph HCLG
|
||||
you need LM in ARPA format and files in table below.
|
||||
|
||||
* Note 1: Building HCLG decoding graph is out of scope this README.
|
||||
* Note 2: Each acoustic model needs corresponding HCLG graph.
|
||||
* Note 3: The phonetic dictionary applied on the vocabulary
|
||||
should always generate only a subset of phones seen in training data!
|
||||
|
||||
=============================== =========================================================================
|
||||
LM.arpa Language model in ARPA format [You should supply it]
|
||||
vocabulary.txt List of words you want to decode [You should supply it]
|
||||
OOV_SYMBOL String representing out of vocabulary word. [You should supply it]
|
||||
dictionary.txt Phonetic dictionary. [You should supply it]
|
||||
exp/tri2b_bmmi/final.mdl Trained acoustic model (AM).
|
||||
exp/tri2b_bmmi/final.tree Phonetic decision tree.
|
||||
=============================== =========================================================================
|
||||
|
|
@ -0,0 +1 @@
|
|||
../vystadial_cz/online_demo/
|
||||
|
|
@ -0,0 +1,22 @@
|
|||
# "queue.pl" uses qsub. The options to it are
|
||||
# options to qsub. If you have GridEngine installed,
|
||||
# change this to a queue you have access to.
|
||||
# Otherwise, use "run.pl", which will run jobs locally
|
||||
# (make sure your --num-jobs options are no more than
|
||||
# the number of cpus on your machine.
|
||||
|
||||
#export train_cmd="queue.pl -q all.q@a*.clsp.jhu.edu"
|
||||
#export decode_cmd="queue.pl -q all.q@a*.clsp.jhu.edu"
|
||||
# export train_cmd="queue.pl -l mf=5g"
|
||||
# export decode_cmd="queue.pl -l mf=5g"
|
||||
export train_cmd="queue.pl -l arch=*64*"
|
||||
export decode_cmd="queue.pl -l arch=*64*"
|
||||
|
||||
# The number of parallel jobs to be started for some parts of the recipe
|
||||
# Make sure you have enough resources(CPUs and RAM) to accomodate this number of jobs
|
||||
njobs=20
|
||||
|
||||
# If you have no GridEngine you can do:
|
||||
#export train_cmd=run.pl
|
||||
#export decode_cmd=run.pl
|
||||
#njobs=2
|
||||
|
|
@ -0,0 +1,254 @@
|
|||
A.M. EY EH M
|
||||
AQUIRE AE K W AY R
|
||||
ABOUTS AH B AW T S
|
||||
ADC AE D K
|
||||
ADDEN AE D EH N
|
||||
ADDENBROOKE AE D EH N B R AH K AH
|
||||
ADDENBROOKE'S AE D EH N B R AH K AH Z
|
||||
AHAR AE AH R
|
||||
ALIMENTUM AE L AH M EH N T AH M
|
||||
ALLENBELL AO L AH N B EH L
|
||||
ALPHA-MILTON AE L F AH M IH L T AH N
|
||||
APOLOGISE AH P AA L AH JH AY Z
|
||||
ARBURY AA R B Y AH R IY
|
||||
ASWELL AH S W EH L
|
||||
AUNTIE'S AE N T IY Z
|
||||
AYLESBRAY EY L EH Z R EY IY
|
||||
BACKSTREET B AE K S T R IY T
|
||||
BADMINTONCOURT B AE D M IH N T AH N K OW R T
|
||||
BEACONSFIELD B IY K AH N Z F IY L D
|
||||
BENE'T B AH N EH T
|
||||
BENET B EH N AH T
|
||||
BENET(2) B AH N EY
|
||||
BENETS B IH N EH T S
|
||||
BENNY'S B EH N IY Z
|
||||
BINH B IH N
|
||||
BLD'S B IY EH L D IY S
|
||||
BLU B L UW
|
||||
BOCHKA B AA CH AH
|
||||
BOCHKAS B AA CH AH Z
|
||||
BOLCHOI B OW L CH OY
|
||||
BOSPHORUS B AA S AH R AH S
|
||||
BROOKFIELDS B R UH K F IY L D Z
|
||||
BUDGENS B AH JH AH N Z
|
||||
BURRELL'S B ER EH L Z
|
||||
BURWASH B AH R W AH SH
|
||||
CAFE-BAR K AH F EY B AA R
|
||||
CAIPIRINHA K EY P AH R IH N HH AH
|
||||
CAIPIRINHAS K EY P AH R IH N HH AH Z
|
||||
CAIUS K EY AH S
|
||||
CAMBOURNE K AE M B OW R N
|
||||
CARLUCCIO'S K AA R L AH K S IY AH Z
|
||||
CARNABY K AA R N AE B IY
|
||||
CARRINGTONS K AA R IH NG T AH N Z
|
||||
CASTLEHILL K AE S AH L HH IH L
|
||||
CATHARINE'S K AE TH R IH N Z
|
||||
CAZIMIR K AE Z AH M AH R
|
||||
CERTIFICATED S ER T IH F IH K AH T IH D
|
||||
CHEAPISH CH IY P IH SH
|
||||
CHERSTON CH EH R S T AH N
|
||||
CHESTERON CH EH S T ER R AH N
|
||||
CHERRYHINTON CH EH R IY HH IH N T AH N
|
||||
CHERRYWOOD CH EH R IY W UH D
|
||||
CHIQUITO CH IH K W AH T OW
|
||||
CHIQUITAS CH IH K W AH T AH Z
|
||||
CINEWORLD S IH N Y UW AH R L D
|
||||
CITYROOMZ S IH T AH R UW M Z
|
||||
COBBLES K AA B AH L Z
|
||||
COLADA K OW L AE D AH
|
||||
COLDHAM K OW L D HH AE M
|
||||
COLDHAM'S K OW L D HH AE M Z
|
||||
COLDHAMS K OW L D HH AE M Z
|
||||
COLOURS K AH L ER Z
|
||||
COMPETITVE K AA M P EH T AH T V
|
||||
COMPETITVE K AA M P EH T AH T V
|
||||
CRICKETERS K R IH K AH T ER Z
|
||||
CROSSWAYS K R AO S W EY Z
|
||||
CUCINA K Y UW S AH N AH
|
||||
CURY K Y UH R IY
|
||||
D'ARRY'S D AE R AH Z
|
||||
DOJO D OW JH OW
|
||||
EFES IY F S
|
||||
EMBARASSING IH M B AA R AE S IH NG
|
||||
ENSUITE EH N S W IY T
|
||||
ENSUITES EH N S W IY T S
|
||||
ERAINA AH R EY N AH
|
||||
ERM AH R M
|
||||
EUROS Y UW R OW Z
|
||||
FAMISHED F AE M IH SH T
|
||||
FEN F EH N
|
||||
FENDITTON F EH N D IH T AH N
|
||||
FENNER'S F EH N ER Z
|
||||
FITZBILLIES F IH T S B IH L AY Z
|
||||
FIZZY F IH Z IY
|
||||
FOODWELL F UW D W EH L
|
||||
FUSIONS F Y UW ZH AH N Z
|
||||
FREUNDE F R UW N D
|
||||
FYNE F AY N
|
||||
GARONDE G AE R AH N D
|
||||
GELDART JH EH L D AA R T
|
||||
GIRTONS G ER1 T AH0 N Z
|
||||
GOG G AA G
|
||||
GONVILLE G AA N V AH L
|
||||
GPS JH IY P IY S
|
||||
GRANTA G R AE N T AH
|
||||
GUINNESSES G UW N EH S AH Z
|
||||
GWYDIR G W IH D AH R
|
||||
GWYDIR G W IH D AH R
|
||||
HAHA HH AE HH AH
|
||||
HAKKA HH AE K AH
|
||||
HALLO HH AE L OW
|
||||
HAUXTON HH AO K S T AH N
|
||||
HEDGE'S HH EH JH IH Z
|
||||
HENRYS HH EH N R IY Z
|
||||
HILARYS HH AY L ER IY Z
|
||||
HINCHINBROOK HH IH NG K IH N B R AH K
|
||||
HIPPY HH IH P IY
|
||||
HISTON HH IH S T AH N
|
||||
HIYA HH AY AY AH
|
||||
HOBSONS HH AA B S AH N Z
|
||||
HOMERTON HH OW M ER T AH N
|
||||
HOPBINE HH AA P B AY N
|
||||
HOTPOT HH AA T P AA T
|
||||
HOUNSLOW HH AW N S L OW
|
||||
HSBC EY CH EH S B IY K
|
||||
HUMBERSTONE HH AH M B IY ER S T OW N
|
||||
ITEMISED AY T AH M IY Z D
|
||||
JASONS JH AE S AH N Z
|
||||
JASONVILLE JH AH S AA N V AH L
|
||||
JEDBURGH JH EH D AH R
|
||||
JINLING JH AY N L AH NG
|
||||
KETTLE'S K EH T AH L Z
|
||||
KILOMETRE K IH L AO M AH T ER
|
||||
KILOMETRE(2) K IH L AH M IY T ER
|
||||
KILOMETRE K IH L AO M AH T ER
|
||||
KILOMETRE(2) K IH L AH M IY T ER
|
||||
KILOMETRE K IH L AO M AH T ER
|
||||
KILOMETRE(2) K IH L AH M IY T ER
|
||||
KILOMETRES K IH L AO M AH T ER Z
|
||||
KILOMETRES(2) K IH L AH M IY T ER Z
|
||||
KINGSHEDGES K IH NG Z HH EH JH IH Z
|
||||
KINK K IH NG K
|
||||
KOHINOOR K AH HH IH N OW AH R
|
||||
KYMMOY K IH M OW IY
|
||||
LABELLING L EY B AH L IH NG
|
||||
LABELLING L EY B AH L IH NG
|
||||
LAGERS L AA G ER Z
|
||||
LENSFIELD L EH N Z F IY L D
|
||||
LIVINGSTONES L IH V IH NG S T OW N Z
|
||||
MAGOG M AE G AH G
|
||||
MAHARAJAH M AH HH AE R AH JH AH
|
||||
MAINCOURSE M EY N K OW R S
|
||||
MARGHERITA M AH R EH R AH T AH
|
||||
MASSARO'S M AH S AA R OW Z
|
||||
MASSIMO'S M AE S IH M OW Z
|
||||
MATHS M AE TH S
|
||||
MAYPOLE M EY P OW L
|
||||
MEGHNA M EH N AH
|
||||
MERCERS M ER S ER Z
|
||||
METRES M IY T ER Z
|
||||
METROPOLE M EH T R OW P OW L
|
||||
METROPOLE M EH T R OW P OW L
|
||||
MEZE M IY Z
|
||||
MICHAELHOUSE M IH CH IY L HH AW Z
|
||||
MID-PRICED M IH D P R AY S T
|
||||
MIDDLEEASTERN M IH D AH L IY S T ER N
|
||||
MITCHUM'S M IH CH AH M Z
|
||||
MULTISTOREY M AH L T AY S T OW R IY
|
||||
NANDO'S N AE N D OW Z
|
||||
NANDOS N AE N D AH Z
|
||||
NAZ N AE Z
|
||||
NEATH N IY TH
|
||||
NEWCHESTERTON N UW CH EH S T ER T AH N
|
||||
NEWNHAM N Y UW N HH AH M
|
||||
NHS EH N EY CH S
|
||||
NU-MEGA N UW M EH G AH
|
||||
NUSHA N AH SH AH
|
||||
P.M. P IY EH M
|
||||
PANAHAR P AE N AH AH R
|
||||
PASTY P EY S T IY
|
||||
PATISSERIE P EY T IY Z ER IY
|
||||
PATISSIER P EY T IY Z IY ER
|
||||
PEKINGS P IY1 K IH1 NG Z
|
||||
PERNE P AH R N IY
|
||||
PICTUREHOUSE P IH K CH ER HH AW S
|
||||
PIPASHA P IH P AH SH AH
|
||||
PLAINING P L EY N IH NG
|
||||
PLC P IY EH L K
|
||||
POLONIA P AH L OW N Y AH
|
||||
PREZZO P R IY Z OW
|
||||
PRICERANGE P R AY S R EY N JH
|
||||
PUNTING P AH N T IH NG
|
||||
QUAYSIDE K W EY S AY D
|
||||
QUEENS' K W IY N Z
|
||||
RADEGUND R AE D AH G AH N D
|
||||
ROMWELL R AA M W EH L
|
||||
ROSCO R AO S K OW
|
||||
REALISED R IY L IY Z D
|
||||
RECOGNISE R IH K AH N AY Z
|
||||
RECOGNISED R AH K AA G N AH S T
|
||||
RECOGNISING R IH K AH N AY Z AH NG
|
||||
ROMSEY R AA M Z IY
|
||||
ROSEFORD R OW Z F OW R D
|
||||
RUGBYFIELD R AH G B AY F IY L D
|
||||
SAINSBURY'S S EY N S B EH R IY Z
|
||||
SAVINO'S S AA V IY N OW Z
|
||||
SAVOURY S AH V AW R IY
|
||||
SECS S EH K S
|
||||
SERGU S AH R G Y UW
|
||||
SHADDAI SH AE D EY
|
||||
SHELFORD SH EH L F AH R D
|
||||
SHIRAZ SH IH R AH Z
|
||||
SIDGWICK S IH JH W AH K
|
||||
SLEEPERZ S L IY P AH R Z
|
||||
SNAKATAK S N AE K AH T AH K
|
||||
SORRENTO S AO R EH N T OW
|
||||
SPELT S P EH L T
|
||||
SQUASHCOURT S K W AA SH K OW R T
|
||||
ST. S EY N T
|
||||
STANSTED S T AE N S T AH D
|
||||
STAZIONE S T EY Z IY OW N
|
||||
STOREY'S S T AO R IY Z
|
||||
STOUTS S T AW T S
|
||||
STOW-CUM-QUY S T OW K AH M K W IY
|
||||
STURTON S T AH R T AH N
|
||||
SWIMMINGPOOL S W IH M IH NG P UW L
|
||||
TAKEAWAY T EY K AH W EY
|
||||
TANDOORI T AE N D AH R IY
|
||||
TARIF T AE R AH F
|
||||
TATTIES T AE T IY Z
|
||||
TENISON T EH N IY Z AO N
|
||||
TEXMEX T EH K S M EH K S
|
||||
TEQUILAS T EH K W AH L AH Z
|
||||
TERI-AKI T EH R IY AA K IY
|
||||
THANKYOU TH AE NG K Y UW
|
||||
THATS DH AE T S
|
||||
THODAY TH AA D EY IY
|
||||
THOROUGHBREADS TH ER OW B R EH D Z
|
||||
TIMINGS T AY M IH NG Z
|
||||
TOURISTY T UH R IH S T IY
|
||||
TRAVELLER'S T R AE V AH L ER Z
|
||||
TROCKEL T R AA K AH L
|
||||
TRUMPINGTON T R AH M P IH NG T AH N
|
||||
ULMANN AH L M AH N
|
||||
UPTO AH P T OW
|
||||
VENUE'S V EH N Y UW Z
|
||||
WAGAMAMA W AH G AE M AH M AH
|
||||
WARKWORTH W AO R K W AH R TH
|
||||
WATERBEACH W AO T ER B IY CH
|
||||
WI-FI W IY F AY
|
||||
WIFI W IY F AY
|
||||
WHEATSHEAF W IY T SH IY F
|
||||
WHEREABOUT W IY R AH B AW T
|
||||
WOLLASTON W AA L AE S T AH N
|
||||
ZIZZI Z IH Z IY
|
||||
BARNOBUS B AA R N AH B AH S
|
||||
MARTIN'S M AA R T AH N Z
|
||||
ROUTE R UW T
|
||||
ROUTE(2) R AW T
|
||||
ROUTES R UW T S
|
||||
ROUTES(2) R AW T S
|
||||
ROUTES(3) R UH T S
|
||||
TERRANIAN T EH R EY N IY AH N
|
||||
YIPEE Y IY P IY
|
||||
UK Y K EY
|
||||
|
|
@ -0,0 +1,3 @@
|
|||
beam=12.0
|
||||
latbeam=6.0
|
||||
max_active=14000
|
||||
|
|
@ -0,0 +1,7 @@
|
|||
# --use-energy=false # non default option. False -> Use C0 instead energy
|
||||
# NUMCEPS in HTK is without C0: there 12 here 13 - default
|
||||
--low-freq=125
|
||||
--high-freq=3800
|
||||
# --htk-compat
|
||||
# --remove-dc-offset # equivalent ZMEANSOURCE in HTK
|
||||
# --subtract-mean # not recommended to do it this way
|
||||
|
|
@ -0,0 +1,42 @@
|
|||
#!/bin/bash
|
||||
|
||||
# EVERY_N utterance is used for training
|
||||
# EVERY_N=3 -> we use one third of training data
|
||||
export EVERY_N=1
|
||||
export TEST_SETS="dev test"
|
||||
|
||||
# Directories set up
|
||||
export DATA_ROOT=`pwd`/data # expects subdirectories train + $TEST_SETS
|
||||
export WORK=`pwd`/lang_prep
|
||||
export EXP=`pwd`/exp
|
||||
export TGT_MODELS=trained_models
|
||||
|
||||
# Specify paths to arpa models. Paths may not contain spaces.
|
||||
# Specify build0 or build1 or build2, .. for building (zero|uni|bi)-gram LM.
|
||||
# Note: The LM file name should not contain underscore "_"!
|
||||
# Otherwise the results will be reported without the LM with underscore."
|
||||
export LMs="build0 build2"
|
||||
|
||||
# Use path to prebuilt dictionary or 'build' command in order to build dictionary
|
||||
# export DICTIONARY="../../resources/lm/caminfo/dict"
|
||||
export DICTIONARY="build"
|
||||
|
||||
|
||||
# Borders for estimating LM model weight.
|
||||
# LMW is tuned on development set and applied on test set.
|
||||
export min_lmw=9
|
||||
export max_lmw=20
|
||||
|
||||
# Number of states for phonem training
|
||||
export pdf=1200
|
||||
|
||||
# Maximum number of Gaussians used for training
|
||||
export gauss=19200
|
||||
|
||||
export train_mmi_boost=0.05
|
||||
|
||||
export mmi_beam=16.0
|
||||
export mmi_lat_beam=10.0
|
||||
|
||||
# --fake -> NO CMVN; empty -> CMVN (pykaldi decoders can not handle CMVN -> fake)
|
||||
export fake="--fake"
|
||||
|
|
@ -0,0 +1,41 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
expdir=$1; shift
|
||||
tgtdir=$2; shift
|
||||
|
||||
tgtdir="$tgtdir/$name"
|
||||
date="`date +%F_%T.%N`"
|
||||
|
||||
if [[ -d $tgtdir || -f $tgtdir ]] ; then
|
||||
tgtdir="$tgtdir/backup_$date"
|
||||
fi
|
||||
|
||||
|
||||
# This is EXAMPLE SCRIPT you are ENCOURAGED TO CHANGE IT!
|
||||
|
||||
mkdir -p "$tgtdir"
|
||||
cp -rf $expdir "$tgtdir"
|
||||
|
||||
# Collect the results
|
||||
|
||||
local/results.py $EXP > "$tgtdir"/results.log
|
||||
echo "Date: $date" >> "$tgtdir"/results.log
|
||||
size=`du -hs "$tgtdir"`
|
||||
echo "Size of backup: $size" >> "$tgtdir"/results.log
|
||||
|
||||
echo; echo "DATA successfully copied to $tgtdir"; echo
|
||||
|
|
@ -0,0 +1,68 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright 2012 Vassil Panayotov
|
||||
# 2013 Ondrej Platek
|
||||
# Apache 2.0
|
||||
|
||||
echo "===test_sets Formating data ..."
|
||||
langdir=$1; shift
|
||||
LMs=$1; shift
|
||||
lmdir=$1; shift
|
||||
lexicon=$1; shift
|
||||
|
||||
# Next, for each type of language model, create the corresponding FST
|
||||
# and the corresponding lang_test_* directory.
|
||||
for lm in $LMs ; do
|
||||
tgt=${langdir}_`basename "$lm"`
|
||||
lmp=$lmdir/`basename $lm`
|
||||
|
||||
tmpdir=$tgt/tmp
|
||||
mkdir -p $tgt
|
||||
mkdir -p $tmpdir
|
||||
|
||||
echo "--- Preparing the grammar transducer (G.fst) from $lmp in $tgt ..."
|
||||
|
||||
for f in phones.txt words.txt phones.txt L.fst L_disambig.fst phones ; do
|
||||
ln -s $langdir/$f $tgt/$f 2> /dev/null
|
||||
done
|
||||
|
||||
cat $lmp | utils/find_arpa_oovs.pl $tgt/words.txt > $tmpdir/oovs.txt
|
||||
|
||||
# grep -v '<s> <s>' because the LM seems to have some strange and useless
|
||||
# stuff in it with multiple <s>'s in the history. Encountered some other similar
|
||||
# things in a LM from Geoff. Removing all "illegal" combinations of <s> and </s>,
|
||||
# which are supposed to occur only at being/end of utt. These can cause
|
||||
# determinization failures of CLG [ends up being epsilon cycles].
|
||||
|
||||
cat $lmp | \
|
||||
grep -v '<s> <s>\|</s> <s>\|</s> </s>' | \
|
||||
arpa2fst - | fstprint | \
|
||||
utils/remove_oovs.pl $tmpdir/oovs.txt | \
|
||||
utils/eps2disambig.pl | utils/s2eps.pl | fstcompile --isymbols=$tgt/words.txt \
|
||||
--osymbols=$tgt/words.txt --keep_isymbols=false --keep_osymbols=false | \
|
||||
fstrmepsilon > $tgt/G.fst
|
||||
fstisstochastic $tgt/G.fst
|
||||
# The output is like:
|
||||
# 9.14233e-05 -0.259833
|
||||
# we do expect the first of these 2 numbers to be close to zero (the second is
|
||||
# nonzero because the backoff weights make the states sum to >1).
|
||||
# Because of the <s> fiasco for these particular LMs, the first number is not
|
||||
# as close to zero as it could be.
|
||||
|
||||
# Everything below is only for diagnostic.
|
||||
# Checking that G has no cycles with empty words on them (e.g. <s>, </s>);
|
||||
# this might cause determinization failure of CLG.
|
||||
# #0 is treated as an empty word.
|
||||
mkdir -p $tmpdir/g
|
||||
awk '{if(NF==1){ printf("0 0 %s %s\n", $1,$1); }} END{print "0 0 #0 #0"; print "0";}' \
|
||||
< "$lexicon" >$tmpdir/g/select_empty.fst.txt
|
||||
fstcompile --isymbols=$tgt/words.txt --osymbols=$tgt/words.txt \
|
||||
$tmpdir/g/select_empty.fst.txt | \
|
||||
fstarcsort --sort_type=olabel | fstcompose - $tgt/G.fst > $tmpdir/g/empty_words.fst
|
||||
fstinfo $tmpdir/g/empty_words.fst | grep cyclic | grep -w 'y' &&
|
||||
echo "Language model has cycles with empty words" && exit 1
|
||||
|
||||
# rm -rf $tmpdir # TODO debugging
|
||||
echo "*** Succeeded in creating G.fst for $tgt"
|
||||
|
||||
done # for lm in $LMs ; do
|
||||
|
|
@ -0,0 +1,89 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
locdata=$1; shift
|
||||
train_text=$1; shift
|
||||
test_text=$1; shift
|
||||
local_lm=$1; shift
|
||||
lms=$1; shift
|
||||
|
||||
|
||||
mkdir -p $local_lm
|
||||
|
||||
echo "=== Preparing the LM ..."
|
||||
|
||||
function build_0gram {
|
||||
echo "=== Building zerogram $lm from ${transcr}. ..."
|
||||
transcr=$1; lm=$2
|
||||
cut -d' ' -f2- $transcr | tr ' ' '\n' | sort -u > $lm
|
||||
echo "<s>" >> $lm
|
||||
echo "</s>" >> $lm
|
||||
python -c """
|
||||
import math
|
||||
with open('$lm', 'r+') as f:
|
||||
lines = f.readlines()
|
||||
p = math.log10(1/float(len(lines)));
|
||||
lines = ['%f\\t%s'%(p,l) for l in lines]
|
||||
f.seek(0); f.write('\\n\\\\data\\\\\\nngram 1= %d\\n\\n\\\\1-grams:\\n' % len(lines))
|
||||
f.write(''.join(lines) + '\\\\end\\\\')
|
||||
"""
|
||||
}
|
||||
|
||||
for lm in $lms ; do
|
||||
lm_base=`basename $lm`
|
||||
if [ ${lm_base%[0-6]} != 'build' ] ; then
|
||||
cp $lm $local_lm
|
||||
else
|
||||
# We will build the LM 'build[0-9].arpa
|
||||
lm_order=${lm_base#build}
|
||||
|
||||
echo "=== Building LM of order ${lm_order}..."
|
||||
if [ $lm_order -eq 0 ] ; then
|
||||
echo "Zerogram $lm_base LM is build from text: $test_text"
|
||||
cut -d' ' -f2- $test_text | sed -e 's:^:<s> :' -e 's:$: </s>:' | \
|
||||
sort -u > $locdata/lm_test.txt
|
||||
build_0gram $locdata/lm_test.txt $local_lm/${lm_base}
|
||||
else
|
||||
echo "LM $lm_base is build from text: $train_text"
|
||||
cut -d' ' -f2- $train_text | sed -e 's:^:<s> :' -e 's:$: </s>:' | \
|
||||
sort -u > $locdata/lm_train.txt
|
||||
ngram-count -text $locdata/lm_train.txt -order ${lm_order} \
|
||||
-wbdiscount -interpolate -lm $local_lm/${lm_base}
|
||||
fi
|
||||
fi
|
||||
done
|
||||
echo "*** LMs preparation finished!"
|
||||
|
||||
echo "=== Preparing the vocabulary ..."
|
||||
|
||||
if [ "$DICTIONARY" == "build" ]; then
|
||||
echo; echo "Building dictionary from train data"; echo
|
||||
cut -d' ' -f2- $train_text | tr ' ' '\n' > $locdata/vocab-full-raw.txt
|
||||
else
|
||||
echo; echo "Using predefined dictionary: ${DICTIONARY}"
|
||||
echo "Throwing away first 2 rows."; echo
|
||||
tail -n +3 $DICTIONARY | cut -f 1 > $locdata/vocab-full-raw.txt
|
||||
fi
|
||||
|
||||
echo '</s>' >> $locdata/vocab-full-raw.txt
|
||||
echo "Removing from vocabulary _NOISE_, and all '_' words from vocab-full.txt"
|
||||
cat $locdata/vocab-full-raw.txt | grep -v '_' | \
|
||||
sort -u > $locdata/vocab-full.txt
|
||||
echo "*** Vocabulary preparation finished!"
|
||||
|
||||
|
||||
echo "Removing from vocabulary _NOISE_, and all '_' words from vocab-test.txt"
|
||||
cut -d' ' -f2 $test_text | tr ' ' '\n' | grep -v '_' | sort -u > $locdata/vocab-test.txt
|
||||
|
||||
|
|
@ -0,0 +1,54 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
# The vystadial data are specific by having following marks in transcriptions
|
||||
# _INHALE_
|
||||
# _LAUGH_
|
||||
# _EHM_HMM_
|
||||
# _NOISE_
|
||||
# _EHM_HMM_
|
||||
# _SIL_
|
||||
|
||||
locdict=$1; shift
|
||||
|
||||
echo "--- Prepare nonsilence phone lists ..."
|
||||
# We suppose only nonsilence_phones in lexicon right now
|
||||
awk '{for(n=2;n<=NF;n++) { p[$n]=1; }} END{for(x in p) {print x}}' \
|
||||
$locdict/lexicon.txt | sort > $locdict/nonsilence_phones.txt
|
||||
|
||||
echo "--- Adding silence phones to lexicon ..."
|
||||
echo "_SIL_ SIL" >> $locdict/lexicon.txt
|
||||
echo "_EHM_HMM_ EHM" >> $locdict/lexicon.txt
|
||||
echo "_INHALE_ INH" >> $locdict/lexicon.txt
|
||||
echo "_LAUGH_ LAU" >> $locdict/lexicon.txt
|
||||
echo "_NOISE_ NOI" >> $locdict/lexicon.txt
|
||||
|
||||
echo "--- Sorting lexicon in place..."
|
||||
sort $locdict/lexicon.txt -o $locdict/lexicon.txt
|
||||
|
||||
echo "--- Prepare silence phone lists ..."
|
||||
echo SIL > $locdict/silence_phones.txt
|
||||
echo EHM >> $locdict/silence_phones.txt
|
||||
echo INH >> $locdict/silence_phones.txt
|
||||
echo LAU >> $locdict/silence_phones.txt
|
||||
echo NOI >> $locdict/silence_phones.txt
|
||||
|
||||
echo SIL > $locdict/optional_silence.txt
|
||||
|
||||
# Some downstream scripts expect this file exists, even if empty
|
||||
touch $locdict/extra_questions.txt
|
||||
|
||||
echo "*** Creating phone lists finished!"
|
||||
|
|
@ -0,0 +1,18 @@
|
|||
#!/bin/bash
|
||||
# example usage:
|
||||
# ./local/create_sample.sh /ha/projects/vystadial/data/asr/en/voip/ Results/vystadial-sample/ test 100
|
||||
# note that it suppose there are only *.wav and *.wav.trn and the
|
||||
# the n is the numbero of files in the directory
|
||||
|
||||
src=$1
|
||||
tgt=$2
|
||||
typ=$3 # dev/test/train
|
||||
n=$4
|
||||
|
||||
src_dir=$src/$typ
|
||||
tgt_dir=$tgt/$typ
|
||||
mkdir -p $tgt_dir
|
||||
ls $src_dir | head -n $n \
|
||||
| while read f ; do
|
||||
cp $src_dir/$f $tgt_dir
|
||||
done
|
||||
|
|
@ -0,0 +1,108 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
#
|
||||
# Makes train/test splits
|
||||
# local/voxforge_data_prep.sh --nspk_test ${nspk_test} ${SELECTED} || exit 1
|
||||
# create files: (TYPE=train|test)
|
||||
# a) ${TYPE}_trans.txt: ID transcription capitalized! No interputction
|
||||
# b) ${TYPE}_wav.scp: ID path2ID.wav
|
||||
# c) $TYPE.utt2spk: ID-recording ID-speaker
|
||||
# s) $TYPE.spk2utt
|
||||
# e) $TYPE.spk2gender all speakers are male
|
||||
# we have ID-recording = ID-speaker
|
||||
|
||||
# The vystadial data are specific by having following marks in transcriptions
|
||||
# _INHALE_
|
||||
# _LAUGH_
|
||||
# _EHM_HMM_
|
||||
# _NOISE_
|
||||
# _EHM_HMM_
|
||||
# _SIL_
|
||||
|
||||
# renice 20 $$
|
||||
|
||||
every_n=1
|
||||
|
||||
[ -f path.sh ] && . ./path.sh # source the path.
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
|
||||
if [ $# -ne 4 ] ; then
|
||||
echo "Usage: local/data_split.sh [--every-n 30] <data-directory> <local-directory> <LMs> <Test-Sets> <tgt-dir>";
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
DATA=$1; shift
|
||||
locdata=$1; shift
|
||||
LMs=$1; shift
|
||||
test_sets=$1; shift
|
||||
tgt_dir=$1; shift
|
||||
|
||||
echo "LMs $LMs test_sets $test_sets"
|
||||
|
||||
|
||||
echo "=== Starting initial Vystadial data preparation ..."
|
||||
echo "--- Making test/train data split from $DATA taking every $every_n recording..."
|
||||
|
||||
mkdir -p $locdata
|
||||
|
||||
i=0
|
||||
for s in $test_sets train ; do
|
||||
mkdir -p $locdata/$s
|
||||
ls $DATA/$s/ | sed -n /.*wav$/p |\
|
||||
while read wav ; do
|
||||
((i++)) # bash specific
|
||||
if [[ $i -ge $every_n ]] ; then
|
||||
i=0
|
||||
pwav=$DATA/$s/$wav
|
||||
trn=`cat $DATA/$s/$wav.trn`
|
||||
echo "$wav $pwav" >> $locdata/$s/wav.scp
|
||||
echo "$wav $wav" >> $locdata/$s/utt2spk
|
||||
echo "$wav $wav" >> $locdata/$s/spk2utt
|
||||
echo "$wav $trn" >> $locdata/$s/trans.txt
|
||||
# Ignoring gender -> label all recordings as male
|
||||
echo "$wav M" >> $locdata/spk2gender
|
||||
fi
|
||||
done # while read wav
|
||||
|
||||
for f in wav.scp utt2spk spk2utt trans.txt ; do
|
||||
sort "$locdata/$s/$f" -k1 -u -o "$locdata/$s/$f" # sort in place
|
||||
done # for f
|
||||
|
||||
done # for in $test_sets train
|
||||
|
||||
echo "Set 1:1 relation for spk2utt: spk in $test_sets AND train, sort in place"
|
||||
sort "$locdata/spk2gender" -k1 -o "$locdata/spk2gender"
|
||||
|
||||
echo "--- Distributing the file lists to train and ($test_sets x $LMs) directories ..."
|
||||
mkdir -p $WORK/train
|
||||
cp $locdata/train/wav.scp $WORK/train/wav.scp || exit 1;
|
||||
cp $locdata/train/trans.txt $WORK/train/text || exit 1;
|
||||
cp $locdata/train/spk2utt $WORK/train/spk2utt || exit 1;
|
||||
cp $locdata/train/utt2spk $WORK/train/utt2spk || exit 1;
|
||||
utils/filter_scp.pl $WORK/train/spk2utt $locdata/spk2gender > $WORK/train/spk2gender || exit 1;
|
||||
|
||||
for s in $test_sets ; do
|
||||
for lm in $LMs; do
|
||||
tgt_dir=$WORK/${s}_`basename ${lm}`
|
||||
mkdir -p $tgt_dir
|
||||
cp $locdata/${s}/wav.scp $tgt_dir/wav.scp || exit 1;
|
||||
cp $locdata/${s}/trans.txt $tgt_dir/text || exit 1;
|
||||
cp $locdata/${s}/spk2utt $tgt_dir/spk2utt || exit 1;
|
||||
cp $locdata/${s}/utt2spk $tgt_dir/utt2spk || exit 1;
|
||||
utils/filter_scp.pl $tgt_dir/spk2utt $locdata/spk2gender > $tgt_dir/spk2gender || exit 1;
|
||||
done
|
||||
done
|
||||
|
|
@ -0,0 +1,46 @@
|
|||
#!/bin/bash
|
||||
# Copyright Ondrej Platek Apache 2.0
|
||||
|
||||
DATA_ROOT=$1
|
||||
|
||||
url=https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11858/00-097C-0000-0023-4671-4/data_voip_en.tgz
|
||||
name=data_voip_en
|
||||
extract_file=205859
|
||||
|
||||
mkdir -p $DATA_ROOT
|
||||
|
||||
if [ ! -f $DATA_ROOT/${name}.tgz ] ; then
|
||||
wget $url -O $DATA_ROOT/${name}.tgz || exit 1
|
||||
echo "Data successfully downloaded"
|
||||
fi
|
||||
|
||||
if [[ ! -d $DATA_ROOT/$name && -e $DATA_ROOT/$name ]] ; then
|
||||
echo "The $DATA_ROOT/$name is not a directory and we cannot extract the data!"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ ! -d $DATA_ROOT/$name ] ; then
|
||||
mkdir $DATA_ROOT/$name
|
||||
tar xfv $DATA_ROOT/${name}.tgz -C $DATA_ROOT | \
|
||||
while read line; do
|
||||
x=$((x+1))
|
||||
echo -en "$x extracted from $extract_file files.\r"
|
||||
done
|
||||
fi
|
||||
|
||||
if [ -d $DATA_ROOT/$name ] ; then
|
||||
echo "Checking if data extracted correctly"
|
||||
num_files=`find $DATA_ROOT/$name -name '*' | wc -l`
|
||||
if [ ! $num_files -eq $extract_file ] ; then
|
||||
echo "Data extraction failed! Extracted $num_files instead of $extract_file"
|
||||
exit 1;
|
||||
fi
|
||||
echo "It seams that data are extracted correctly"
|
||||
fi
|
||||
|
||||
pushd $DATA_ROOT
|
||||
for t in test train dev ; do
|
||||
ln -s $name/$t
|
||||
done
|
||||
ln -s $name/arpa_bigram arpa-bigram
|
||||
popd
|
||||
|
|
@ -0,0 +1,50 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
tgt=$1; shift
|
||||
exp=$1; shift
|
||||
lang=$1; shift
|
||||
|
||||
mkdir -p $tgt
|
||||
|
||||
echo "--- Exporting models to $tgt ..."
|
||||
|
||||
# See local/save_check.sh which saves the settings at the beginning for details
|
||||
# cp -f $exp/alex_gitlog.log $exp/alex_gitdiff.log $exp/experiment_bash_vars.log $tgt
|
||||
cp -f $exp/experiment_bash_vars.log $tgt
|
||||
|
||||
# Store also the results
|
||||
cp -f $exp/results.log $tgt/results.log
|
||||
|
||||
|
||||
cp -f common/mfcc.conf $tgt
|
||||
|
||||
cp -f $exp/tri2a/final.mdl $tgt/tri2a.mdl
|
||||
cp -f $exp/tri2a/tree $tgt/tri2a.tree
|
||||
|
||||
cp -f $exp/tri2b/final.mdl $tgt/tri2b.mdl
|
||||
cp -f $exp/tri2b/tree $tgt/tri2b.tree
|
||||
cp -f $exp/tri2b/final.mat $tgt/tri2b.mat
|
||||
|
||||
cp -f $exp/tri2b_mmi_b*/final.mdl $tgt/tri2b_bmmi.mdl
|
||||
cp -f $exp/tri2b/tree $tgt/tri2b_bmmi.tree
|
||||
cp -f $exp/tri2b_mmi_b*/final.mat $tgt/tri2b_bmmi.mat
|
||||
|
||||
cp -f $lang/phones.txt $lang/phones/silence.csl $tgt
|
||||
|
||||
|
||||
# FIXME do I need splice_opts for something?
|
||||
|
|
@ -0,0 +1,172 @@
|
|||
#!perl -w
|
||||
|
||||
#
|
||||
# ====================================================================
|
||||
# Copyright (C) 1999-2008 Carnegie Mellon University and Alexander
|
||||
# Rudnicky. All rights reserved.
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without
|
||||
# modification, are permitted provided that the following conditions
|
||||
# are met:
|
||||
#
|
||||
# 1. Redistributions of source code must retain the above copyright
|
||||
# notice, this list of conditions and the following disclaimer.
|
||||
#
|
||||
# 2. Redistributions in binary form must reproduce the above copyright
|
||||
# notice, this list of conditions and the following disclaimer in
|
||||
# the documentation and/or other materials provided with the
|
||||
# distribution.
|
||||
#
|
||||
# This work was supported in part by funding from the Defense Advanced
|
||||
# Research Projects Agency, the Office of Naval Research and the National
|
||||
# Science Foundation of the United States of America, and by member
|
||||
# companies of the Carnegie Mellon Sphinx Speech Consortium. We acknowledge
|
||||
# the contributions of many volunteers to the expansion and improvement of
|
||||
# this dictionary.
|
||||
#
|
||||
# THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND
|
||||
# ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
|
||||
# THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
||||
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY
|
||||
# NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
||||
# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
||||
# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
||||
# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
||||
# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
||||
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
#
|
||||
# ====================================================================
|
||||
#
|
||||
|
||||
# [20050309] (air) Created.
|
||||
# strip out stress marks from a cmudict, producing a "SphinxPhones_40" dictionary
|
||||
# [20080420] (air) Changed to pass comments.
|
||||
# Fixed output collation sequence; DOS eol's
|
||||
# [20090309] (air) fixed duplicate pron and collation bugs
|
||||
# [20090331] (air) restored standard collation order (since other stuff deppends on it)
|
||||
# [20090629] (air) do not put comments into SPHINX_40 version; not all software deals
|
||||
# [20100118] (air) added $VERBOSE; this should really be a cmdline flag...
|
||||
#
|
||||
|
||||
|
||||
$VERBOSE = 0;
|
||||
|
||||
my $basecount = 0;
|
||||
my $dupl = 0;
|
||||
my $base = 0;
|
||||
my $varia = 0;
|
||||
|
||||
if ( scalar @ARGV ne 2 ) { die "usage: make_baseform <input> <output>\n"; }
|
||||
|
||||
open(IN, $ARGV[0]) || die "can't open $ARGV[0] for reading!\n";
|
||||
open(OUT,">$ARGV[1]") || die "can't open $ARGV[1] for writing!\n";
|
||||
|
||||
@header = (); # header comment lines (passed through)
|
||||
%dict = (); # words end up in here
|
||||
%histo = (); # some statistics on variants
|
||||
|
||||
get_dict(\%dict,\@header,IN); # process the entries
|
||||
|
||||
# what have we got?
|
||||
print STDERR "$basecount forms processed\n";
|
||||
print STDERR "$base baseforms, $varia variants and $dupl duplicates found.\n";
|
||||
print STDERR "variant distribution:\n";
|
||||
foreach $var ( sort keys %histo ) {
|
||||
print STDERR "$var\t$histo{$var}\n";
|
||||
}
|
||||
|
||||
# print special comments (copyright, etc.)
|
||||
# removed since it messes some things up...
|
||||
# foreach $h (@header) { print OUT "$h\n"; }
|
||||
|
||||
# print out each entry
|
||||
%dict_out = ();
|
||||
foreach $w (sort keys %dict) {
|
||||
$var=1; # variants will number starting with 2
|
||||
foreach $p ( @{$dict{$w}} ) {
|
||||
if ($var eq 1) {
|
||||
$dict_out{$w} = $p;
|
||||
$var++;
|
||||
} else {
|
||||
$dict_out{"$w($var)"} = $p;
|
||||
$var++;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
foreach $entry ( sort keys %dict_out ) {
|
||||
print OUT "$entry\t$dict_out{$entry}\n";
|
||||
}
|
||||
|
||||
close(IN);
|
||||
close(OUT);
|
||||
|
||||
#
|
||||
#
|
||||
# read in a dictionary
|
||||
sub get_dict {
|
||||
my $dict = shift; # data structure with dictionary entries
|
||||
my $header = shift;
|
||||
my $target = shift; # input file handle
|
||||
|
||||
while (<$target>) {
|
||||
s/[\r\n]+$//g; # DOS-robust chomp;
|
||||
|
||||
# process comments; blank lines ignored
|
||||
# presume that ";;; #" will be collected and emitted at the top
|
||||
if ($_ =~ /^;;; \#/) { push @$header, $_; next; } # save header info
|
||||
elsif ( $_ =~ /^;;;/ ) { next; } # ignore plain comments
|
||||
elsif ( $_ =~ /^\s*$/ ) { next; } # ignore blank lines
|
||||
|
||||
# extract the word,pron pair and prepare for processing
|
||||
($word,$pron) = /(.+?)\s+(.+?)$/;
|
||||
if (! defined $word) { print STDERR "bad entry (no head word): $_\n"; next; }
|
||||
|
||||
$basecount++;
|
||||
|
||||
if ($word =~ /\)$/) { # variant
|
||||
($root,$variant) = ($word =~ m/(.+?)\((.+?)\)/);
|
||||
} else {
|
||||
$root = $word;
|
||||
$variant = 0;
|
||||
}
|
||||
$pron = &strip_stress($pron);
|
||||
|
||||
# found a new baseform; set it up
|
||||
if ( ! defined $dict->{$root} ) {
|
||||
$dict->{$root}[0] = $pron;
|
||||
$base++;
|
||||
next;
|
||||
}
|
||||
|
||||
# old baseform; see if, after removed stress, pron is a duplicate
|
||||
foreach $var ( @{$dict->{$root}} ) {
|
||||
if ( $var eq $pron ) {
|
||||
if ($VERBOSE) {print STDERR "duplicate entry: $root ($variant) $pron\n";}
|
||||
$dupl++;
|
||||
$pron = "";
|
||||
last;
|
||||
}
|
||||
}
|
||||
|
||||
# it's a new variant on an existing baseform, keep it
|
||||
if ( $pron ne "" ) {
|
||||
push @{$dict->{$root}}, $pron;
|
||||
$varia++;
|
||||
$histo{scalar @{$dict->{$root}}}++; # track variant stats
|
||||
if ( scalar @{$dict->{$root}} > 4 ) { print STDERR "$root -- ",scalar @{$dict->{$root}},"\n"; }
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
# strip stress marks from phonetic symbols
|
||||
sub strip_stress {
|
||||
@pron = split " ", $_[0];
|
||||
my $p;
|
||||
foreach $p (@pron) { if ( $p =~ /\d$/) { $p =~ s/(\d+)$//; } }
|
||||
return ( join(" ",@pron));
|
||||
}
|
||||
|
||||
#
|
||||
|
|
@ -0,0 +1,35 @@
|
|||
#!/bin/bash
|
||||
|
||||
locdata=$1
|
||||
locdict=$2
|
||||
|
||||
cmu_dict=common/cmudict.0.7a
|
||||
cmu_ext=common/cmudict.ext
|
||||
|
||||
mkdir -p $locdict
|
||||
|
||||
if [ ! -f $cmu_dict ] ; then
|
||||
echo "--- Downloading CMU dictionary ..."
|
||||
svn export http://svn.code.sf.net/p/cmusphinx/code/trunk/cmudict/cmudict.0.7a \
|
||||
$cmu_dict || exit 1;
|
||||
fi
|
||||
|
||||
echo; echo "If common/cmudict.ext exists, add extra pronunciation to dictionary" ; echo
|
||||
cat $cmu_dict $cmu_ext > $locdict/cmudict_ext.txt 2> /dev/null # ignoring if no extension
|
||||
|
||||
echo "--- Striping stress and pronunciation variant markers from cmudict ..."
|
||||
perl local/make_baseform.pl \
|
||||
$locdict/cmudict_ext.txt /dev/stdout |\
|
||||
sed -e 's:^\([^\s(]\+\)([0-9]\+)\(\s\+\)\(.*\):\1\2\3:' > $locdict/cmudict-plain.txt
|
||||
|
||||
echo "--- Searching for OOV words ..."
|
||||
gawk 'NR==FNR{words[$1]; next;} !($1 in words)' \
|
||||
$locdict/cmudict-plain.txt $locdata/vocab-full.txt |\
|
||||
egrep -v '<.?s>' > $locdict/vocab-oov.txt
|
||||
|
||||
gawk 'NR==FNR{words[$1]; next;} ($1 in words)' \
|
||||
$locdata/vocab-full.txt $locdict/cmudict-plain.txt |\
|
||||
egrep -v '<.?s>' > $locdict/lexicon.txt
|
||||
|
||||
wc -l $locdict/vocab-oov.txt
|
||||
wc -l $locdict/lexicon.txt
|
||||
|
|
@ -0,0 +1,187 @@
|
|||
#!/usr/bin/env python
|
||||
# -*- coding: utf-8 -*-
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
import argparse
|
||||
from numpy import mean
|
||||
import glob
|
||||
import sys
|
||||
import sqlite3
|
||||
|
||||
|
||||
def extract_stat(wer_file):
|
||||
wer, ser = None, None
|
||||
try:
|
||||
with open(wer_file, 'r') as f:
|
||||
s = f.readlines()
|
||||
wer = float(s[1].split()[1])
|
||||
ser = float(s[2].split()[1])
|
||||
|
||||
except Exception as e:
|
||||
print sys.stderr, 'Error parsing file %s' % wer_file
|
||||
print sys.stderr, str(e)
|
||||
return wer, ser
|
||||
|
||||
|
||||
def extractResults(path):
|
||||
wer_files = glob.glob('%s/*/decode_*/*wer_*' % path)
|
||||
table = []
|
||||
for wf in wer_files:
|
||||
try:
|
||||
exp, decode_dir, wer_f = wf.split('/')[-3:]
|
||||
# last split: decode_it3_dev_build0 -> (dev, build0)
|
||||
lm = decode_dir.split('_')[-1]
|
||||
dataset = decode_dir.split('_')[-2]
|
||||
lm_w = int(wer_f[4:]) # strip wer_ from wer_19
|
||||
wer, ser = extract_stat(wf)
|
||||
table.append((exp, dataset, lm, lm_w, wer, ser))
|
||||
except Exception as e:
|
||||
print >> sys.stderr, 'failed to parse %s' % wf
|
||||
print >> sys.stderr, str(e)
|
||||
return table
|
||||
|
||||
|
||||
class Table(object):
|
||||
|
||||
def __init__(self, data=[], colnames=[]):
|
||||
self.data = data
|
||||
self.colnames = colnames
|
||||
self.colSep = '\t'
|
||||
self.lineSep = '\n'
|
||||
|
||||
def data2str(self):
|
||||
strdata = []
|
||||
for r in self.data:
|
||||
strdata.append([str(c) for c in r])
|
||||
return strdata
|
||||
|
||||
def __str__(self):
|
||||
sd = self.data2str()
|
||||
colwidth = [len(c) for c in self.colnames]
|
||||
for j in range(len(colwidth)):
|
||||
for r in sd:
|
||||
colwidth[j] = max(colwidth[j], len(r[j]))
|
||||
|
||||
gaps = [m - len(c) for (m, c) in zip(colwidth, self.colnames)]
|
||||
rows = [self.colSep.join(
|
||||
[c + ' ' * gap for c, gap in zip(self.colnames, gaps)])]
|
||||
for r in sd:
|
||||
gaps = [m - len(c) for (m, c) in zip(colwidth, r)]
|
||||
rows.append(
|
||||
self.colSep.join([c + ' ' * d for c, d in zip(r, gaps)]))
|
||||
return self.lineSep.join(rows)
|
||||
|
||||
|
||||
class LatexTable(Table):
|
||||
|
||||
def __init__(self, data=[], colnames=[]):
|
||||
Table.__init__(self, data, colnames)
|
||||
nc = len(colnames)
|
||||
self.header = '\\begin{tabular}{%s}' % ('c' * nc)
|
||||
self.tail = '\\end{tabular}'
|
||||
self.colSep = ' & '
|
||||
self.lineSep = '\\\\ \n'
|
||||
|
||||
def __str__(self):
|
||||
table_s = super(LatexTable, self).__str__()
|
||||
table_s = table_s.replace('_', '\_')
|
||||
return '%s\n%s\n%s\n' % (self.header, table_s, self.tail)
|
||||
|
||||
|
||||
def Table2LatexTable(table):
|
||||
return LatexTable(table.data, table.colnames)
|
||||
|
||||
|
||||
def createSmallTable(r):
|
||||
d = []
|
||||
for k, v in r.iteritems():
|
||||
w, s, r = v
|
||||
if w == []:
|
||||
minw = None
|
||||
else:
|
||||
minw = min(w) # returns tuple if s is list of tuples
|
||||
if s == []:
|
||||
mins = None
|
||||
else:
|
||||
mins = min(s) # returns tuple if s is list of tuples
|
||||
d.append([k, mean(r), minw, mins])
|
||||
t = Table(d, ['exp', 'RT coef', 'WER', 'SER'])
|
||||
return t
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser(
|
||||
description='Parse experiment directory generated by kaldi vystadial recipe and print statistics')
|
||||
|
||||
parser.add_argument('expath', type=str, action='store',
|
||||
help='Path to experiment directory')
|
||||
parser.add_argument('-l', '--latex', default=False, action='store_true',
|
||||
help='Generate also latex format table')
|
||||
args = parser.parse_args()
|
||||
|
||||
raw_d = extractResults(args.expath)
|
||||
|
||||
conn = sqlite3.connect(':memory:')
|
||||
c = conn.cursor()
|
||||
c.execute(
|
||||
'''CREATE TABLE results (exp text, dataset text, lm text, lm_w int, wer float, ser float)''')
|
||||
c.executemany('INSERT INTO results VALUES (?, ?, ?, ?, ?, ?)', raw_d)
|
||||
|
||||
# get all results sorted
|
||||
# c.execute("SELECT * FROM results ORDER BY exp, dataset, lm, lm_w")
|
||||
# d = c.fetchall()
|
||||
# t = Table(data=d, colnames=['exp', 'set', 'lm', 'LMW', 'WER', 'SER'])
|
||||
# print '%s\n==================' % str(t)
|
||||
|
||||
# best experiment
|
||||
# c.execute("SELECT exp, dataset, lm_w, MIN(wer), ser FROM results ORDER BY exp, lm_w, dataset")
|
||||
# d = c.fetchall()
|
||||
# compare dev and test set by picking up the best experiment
|
||||
# c.execute(("SELECT exp, dataset, lm_w, MIN(wer), ser FROM results "
|
||||
# "GROUP BY exp, lm, dataset ORDER BY exp, lm, dataset"))
|
||||
# d = c.fetchall()
|
||||
# t = Table(data=d, colnames=['exp', 'set', 'lm', 'LMW', 'WER', 'SER'])
|
||||
# print '%s\n==================' % str(t)
|
||||
|
||||
# traditional usage of devset
|
||||
dev_set_query = ("SELECT r.exp, r.lm, r.lm_w FROM results AS r "
|
||||
"INNER JOIN ( SELECT dataset, exp, lm, MIN(wer) as min_wer "
|
||||
" FROM results WHERE dataset=? GROUP BY exp, lm) i "
|
||||
"ON r.exp=i.exp AND r.lm=i.lm AND r.dataset=i.dataset AND r.wer <= i.min_wer "
|
||||
)
|
||||
c.execute(dev_set_query, ('dev',))
|
||||
|
||||
min_dev = c.fetchall()
|
||||
|
||||
# remove duplicates: duplicates if equal mimimum wer in dev set
|
||||
min_dev_un = [(e, lm, lmw) for ((e, lm), lmw) in
|
||||
dict([((e, lm), lmw) for e, lm, lmw in min_dev]).items()]
|
||||
# sort according LM -> sort results according experiment & LMs
|
||||
min_dev_un.sort(key=lambda x: (x[1], x[0]))
|
||||
|
||||
# extract corresponding test results to dev set
|
||||
d = []
|
||||
for exp, lm, lm_w in min_dev_un:
|
||||
c.execute(("SELECT * FROM results WHERE "
|
||||
"dataset='test' AND exp=? AND lm=? AND lm_w=?"),
|
||||
(exp, lm, lm_w))
|
||||
x = c.fetchall()
|
||||
assert (len(x) == 1), "One row should be extracted."
|
||||
d.append(x[0])
|
||||
|
||||
t = Table(data=d, colnames=['exp', 'set', 'LM', 'LMW', 'WER', 'SER'])
|
||||
print str(t)
|
||||
if args.latex:
|
||||
print Table2LatexTable(t)
|
||||
|
|
@ -0,0 +1,49 @@
|
|||
#!/bin/bash
|
||||
# Copyright (c) 2013, Ondrej Platek, Ufal MFF UK <oplatek@ufal.mff.cuni.cz>
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License. #
|
||||
|
||||
EXP=$1
|
||||
|
||||
# make sure that the directories exists
|
||||
conflict=""
|
||||
for d in $@ ; do
|
||||
if [ -d $d ] || [ -f $d ] ; then
|
||||
conflict="$conflict $d"
|
||||
fi
|
||||
done
|
||||
|
||||
if [[ ! -z "$conflict" ]] ; then
|
||||
echo "Running new experiment will create following directories."
|
||||
echo "Some of them already exists!"
|
||||
echo ""
|
||||
echo "Existing directories:"
|
||||
for d in $conflict ; do
|
||||
echo " $d"
|
||||
done
|
||||
read -p "Should I delete the conflicting directories NOW y/n?"
|
||||
case $REPLY in
|
||||
[Yy]* ) echo "Deleting $conflict directories"; rm -rf $conflict;;
|
||||
* ) echo 'Keeping conflicting directories and exiting ...'; exit 1;;
|
||||
esac
|
||||
fi
|
||||
|
||||
for d in $@ ; do
|
||||
mkdir -p $d
|
||||
done
|
||||
|
||||
# Save the variables set up
|
||||
(set -o posix ; set ) > $EXP/experiment_bash_vars.log
|
||||
# git log -1 > $EXP/alex_gitlog.log
|
||||
# git diff > $EXP/alex_gitdiff.log
|
||||
|
|
@ -0,0 +1,53 @@
|
|||
#!/bin/bash
|
||||
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey)
|
||||
# 2014 Mff UK, UFAL (modification: Ondrej Platek)
|
||||
# Apache 2.0
|
||||
|
||||
[ -f ./path.sh ] && . ./path.sh
|
||||
|
||||
# begin configuration section.
|
||||
cmd=run.pl
|
||||
min_lmw=9
|
||||
max_lmw=20
|
||||
#end configuration section.
|
||||
|
||||
[ -f ./path.sh ] && . ./path.sh
|
||||
. parse_options.sh || exit 1;
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: local/score.sh [--cmd (run.pl|queue.pl...)] <data-dir> <lang-dir|graph-dir> <decode-dir>"
|
||||
echo " Options:"
|
||||
echo " --cmd (run.pl|queue.pl...) # specify how to run the sub-processes."
|
||||
echo " --min_lmw <int> # minumum LM-weight for lattice rescoring "
|
||||
echo " --max_lmw <int> # maximum LM-weight for lattice rescoring "
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
data=$1
|
||||
lang_or_graph=$2
|
||||
dir=$3
|
||||
|
||||
symtab=$lang_or_graph/words.txt
|
||||
|
||||
for f in $symtab $dir/lat.1.gz $data/text; do
|
||||
[ ! -f $f ] && echo "score.sh: no such file $f" && exit 1;
|
||||
done
|
||||
|
||||
mkdir -p $dir/scoring/log
|
||||
|
||||
cp $data/text $dir/scoring/test.txt
|
||||
|
||||
$cmd LMW=$min_lmw:$max_lmw $dir/scoring/log/best_path.LMW.log \
|
||||
lattice-best-path --lm-scale=LMW --word-symbol-table=$symtab \
|
||||
"ark:gunzip -c $dir/lat.*.gz|" ark,t:$dir/scoring/LMW.tra || exit 1;
|
||||
|
||||
$cmd LMW=$min_lmw:$max_lmw $dir/scoring/log/score.LMW.log \
|
||||
cat $dir/scoring/LMW.tra \| \
|
||||
utils/int2sym.pl -f 2- $symtab \| \
|
||||
compute-wer --text --mode=present \
|
||||
ark:$dir/scoring/test.txt ark,p:- ">&" $dir/wer_LMW || exit 1;
|
||||
|
||||
# Show results
|
||||
for f in $dir/wer_*; do echo $f; egrep '(WER)|(SER)' < $f; done
|
||||
|
||||
exit 0;
|
||||
|
|
@ -0,0 +1,18 @@
|
|||
# Needed for "correct" sorting
|
||||
export LC_ALL=C
|
||||
export KALDI_ROOT=../../..
|
||||
|
||||
# adding Kaldi binaries to path
|
||||
export PATH=$KALDI_ROOT/src/bin:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/irstlm/bin/:$KALDI_ROOT/src/fstbin/:$KALDI_ROOT/src/gmmbin/:$KALDI_ROOT/src/featbin/:$KALDI_ROOT/src/lm/:$KALDI_ROOT/src/sgmmbin/:$KALDI_ROOT/src/sgmm2bin/:$KALDI_ROOT/src/fgmmbin/:$KALDI_ROOT/src/latbin/:$PWD:$PWD/utils:$PWD/steps:$PATH
|
||||
|
||||
|
||||
srilm_bin=$KALDI_ROOT/tools/srilm/bin/
|
||||
if [ ! -e "$srilm_bin" ] ; then
|
||||
echo "SRILM is not installed in $KALDI_ROOT/tools."
|
||||
echo "May not be able to create LMs!"
|
||||
echo "Please go to $KALDI_ROOT/tools and run ./install_srilm.sh"
|
||||
fi
|
||||
srilm_sub_bin=`find "$srilm_bin" -type d`
|
||||
for d in $srilm_sub_bin ; do
|
||||
export PATH=$d:$PATH
|
||||
done
|
||||
|
|
@ -0,0 +1,165 @@
|
|||
#!/bin/bash
|
||||
# Copyright Ondrej Platek Apache 2.0
|
||||
renice 20 $$
|
||||
|
||||
# Load training parameters
|
||||
. ./env_voip_en.sh
|
||||
# Source optional config if exists
|
||||
[ -f env_voip_en_CUSTOM.sh ] && . ./env_voip_en_CUSTOM.sh
|
||||
|
||||
. ./path.sh
|
||||
|
||||
# If you have cluster of machines running GridEngine you may want to
|
||||
# change the train and decode commands in the file below
|
||||
. ./cmd.sh
|
||||
|
||||
#######################################################################
|
||||
# Preparing acoustic features, LMs and helper files #
|
||||
#######################################################################
|
||||
|
||||
echo " Copy the configuration files to $EXP directory."
|
||||
local/save_check.sh $EXP $WORK/* || exit 1;
|
||||
|
||||
local/download_en_data.sh $DATA_ROOT || exit 1;
|
||||
|
||||
local/data_split.sh --every_n $EVERY_N $DATA_ROOT $WORK/local "$LMs" "$TEST_SETS" || exit 1
|
||||
|
||||
local/create_LMs.sh $WORK/local $WORK/local/train/trans.txt \
|
||||
$WORK/local/test/trans.txt $WORK/local/lm "$LMs" || exit 1
|
||||
|
||||
local/prepare_en_transcription.sh $WORK/local $WORK/local/dict || exit 1
|
||||
|
||||
local/create_phone_lists.sh $WORK/local/dict || exit 1
|
||||
|
||||
utils/prepare_lang.sh $WORK/local/dict '_SIL_' $WORK/local/lang $WORK/lang || exit 1
|
||||
|
||||
local/create_G.sh $WORK/lang "$LMs" $WORK/local/lm $WORK/local/dict/lexicon.txt || exit 1
|
||||
|
||||
echo "Create MFCC features and storing them (Could be large)."
|
||||
for s in train $TEST_SETS ; do
|
||||
steps/make_mfcc.sh --mfcc-config common/mfcc.conf --cmd \
|
||||
"$train_cmd" --nj $njobs $WORK/local/$s $EXP/make_mfcc/$s $WORK/mfcc || exit 1;
|
||||
# Note --fake -> NO CMVN
|
||||
steps/compute_cmvn_stats.sh $fake $WORK/local/$s \
|
||||
$EXP/make_mfcc/$s $WORK/mfcc || exit 1;
|
||||
done
|
||||
|
||||
echo "Decoding is done for each pair (TEST_SET x LMs)"
|
||||
echo "Distribute the links to MFCC feats to all LM variations."
|
||||
cp $WORK/local/train/feats.scp $WORK/train/feats.scp
|
||||
cp $WORK/local/train/cmvn.scp $WORK/train/cmvn.scp
|
||||
for s in $TEST_SETS; do
|
||||
for lm in $LMs; do
|
||||
tgt_dir=${s}_`basename "$lm"`
|
||||
echo "cp $WORK/local/$s/feats.scp $WORK/$tgt_dir/feats.scp"
|
||||
cp $WORK/local/$s/feats.scp $WORK/$tgt_dir/feats.scp
|
||||
echo "cp $WORK/local/$s/cmvn.scp $WORK/$tgt_dir/cmvn.scp"
|
||||
cp $WORK/local/$s/cmvn.scp $WORK/$tgt_dir/cmvn.scp
|
||||
done
|
||||
done
|
||||
|
||||
#######################################################################
|
||||
# Training Acoustic Models #
|
||||
#######################################################################
|
||||
|
||||
echo "Train monophone models on full data -> may be wastefull (can be done on subset)"
|
||||
steps/train_mono.sh --nj $njobs --cmd "$train_cmd" $WORK/train $WORK/lang $EXP/mono || exit 1;
|
||||
|
||||
echo "Get alignments from monophone system."
|
||||
steps/align_si.sh --nj $njobs --cmd "$train_cmd" \
|
||||
$WORK/train $WORK/lang $EXP/mono $EXP/mono_ali || exit 1;
|
||||
|
||||
echo "Train tri1 [first triphone pass]"
|
||||
steps/train_deltas.sh --cmd "$train_cmd" \
|
||||
$pdf $gauss $WORK/train $WORK/lang $EXP/mono_ali $EXP/tri1 || exit 1;
|
||||
|
||||
# draw-tree $WORK/lang/phones.txt $EXP/tri1/tree | dot -Tsvg -Gsize=8,10.5 > graph.svg
|
||||
|
||||
echo "Align tri1"
|
||||
steps/align_si.sh --nj $njobs --cmd "$train_cmd" \
|
||||
--use-graphs true $WORK/train $WORK/lang $EXP/tri1 $EXP/tri1_ali || exit 1;
|
||||
|
||||
echo "Train tri2a [delta+delta-deltas]"
|
||||
steps/train_deltas.sh --cmd "$train_cmd" $pdf $gauss \
|
||||
$WORK/train $WORK/lang $EXP/tri1_ali $EXP/tri2a || exit 1;
|
||||
|
||||
echo "Train tri2b [LDA+MLLT]"
|
||||
steps/train_lda_mllt.sh --cmd "$train_cmd" $pdf $gauss \
|
||||
$WORK/train $WORK/lang $EXP/tri1_ali $EXP/tri2b || exit 1;
|
||||
|
||||
echo "Align all data with LDA+MLLT system (tri2b)"
|
||||
steps/align_si.sh --nj $njobs --cmd "$train_cmd" \
|
||||
--use-graphs true $WORK/train $WORK/lang $EXP/tri2b $EXP/tri2b_ali || exit 1;
|
||||
|
||||
echo "Train MMI on top of LDA+MLLT."
|
||||
steps/make_denlats.sh --nj $njobs --cmd "$train_cmd" \
|
||||
--beam $mmi_beam --lattice-beam $mmi_lat_beam \
|
||||
$WORK/train $WORK/lang $EXP/tri2b $EXP/tri2b_denlats || exit 1;
|
||||
steps/train_mmi.sh $WORK/train $WORK/lang $EXP/tri2b_ali $EXP/tri2b_denlats $EXP/tri2b_mmi || exit 1;
|
||||
|
||||
echo "Train MMI on top of LDA+MLLT with boosting. train_mmi_boost is a e.g. 0.05"
|
||||
steps/train_mmi.sh --boost ${train_mmi_boost} $WORK/train $WORK/lang \
|
||||
$EXP/tri2b_ali $EXP/tri2b_denlats $EXP/tri2b_mmi_b${train_mmi_boost} || exit 1;
|
||||
|
||||
echo "Train MPE."
|
||||
steps/train_mpe.sh $WORK/train $WORK/lang $EXP/tri2b_ali $EXP/tri2b_denlats $EXP/tri2b_mpe || exit 1;
|
||||
|
||||
#######################################################################
|
||||
# Building decoding graph #
|
||||
#######################################################################
|
||||
for lm in $LMs ; do
|
||||
lm=`basename "$lm"`
|
||||
utils/mkgraph.sh --mono $WORK/lang_${lm} $EXP/mono $EXP/mono/graph_${lm} || exit 1
|
||||
utils/mkgraph.sh $WORK/lang_${lm} $EXP/tri1 $EXP/tri1/graph_${lm} || exit 1
|
||||
utils/mkgraph.sh $WORK/lang_${lm} $EXP/tri2a $EXP/tri2a/graph_${lm} || exit 1
|
||||
utils/mkgraph.sh $WORK/lang_${lm} $EXP/tri2b $EXP/tri2b/graph_${lm} || exit 1
|
||||
done
|
||||
|
||||
|
||||
#######################################################################
|
||||
# Decoding #
|
||||
#######################################################################
|
||||
for s in $TEST_SETS ; do
|
||||
for lm in $LMs ; do
|
||||
lm=`basename "$lm"`
|
||||
tgt_dir=${s}_`basename "$lm"`
|
||||
echo "Monophone decoding"
|
||||
# Note: steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
# calls the command line once for each test,
|
||||
# and afterwards averages the WERs into (in this case $EXP/mono/decode/)
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/mono/graph_${lm} $WORK/${tgt_dir} $EXP/mono/decode_${tgt_dir}
|
||||
echo "Decode tri1"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri1/graph_${lm} $WORK/$tgt_dir $EXP/tri1/decode_${tgt_dir}
|
||||
echo "Decode tri2a"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2a/graph_${lm} $WORK/$tgt_dir $EXP/tri2a/decode_${tgt_dir}
|
||||
echo "Decode tri2b [LDA+MLLT]"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b/decode_${tgt_dir}
|
||||
# Note: change --iter option to select the best model. 4.mdl == final.mdl
|
||||
echo "Decode MMI on top of LDA+MLLT."
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --iter 4 --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b_mmi/decode_it4_${tgt_dir}
|
||||
echo "Decode MMI on top of LDA+MLLT with boosting. train_mmi_boost is a number e.g. 0.05"
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --iter 4 --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b_mmi_b${train_mmi_boost}/decode_it4_${tgt_dir};
|
||||
echo "Decode MPE."
|
||||
steps/decode.sh --scoring-opts "--min-lmw $min_lmw --max-lmw $max_lmw" \
|
||||
--config common/decode.conf --iter 4 --nj $njobs --cmd "$decode_cmd" \
|
||||
$EXP/tri2b/graph_${lm} $WORK/$tgt_dir $EXP/tri2b_mpe/decode_it4_${tgt_dir} || exit 1;
|
||||
done
|
||||
done
|
||||
|
||||
|
||||
echo "Successfully trained and evaluated all the experiments"
|
||||
local/results.py $EXP | tee $EXP/results.log
|
||||
|
||||
local/export_models.sh $TGT_MODELS $EXP $WORK/lang
|
||||
|
|
@ -0,0 +1 @@
|
|||
../../wsj/s5/steps
|
||||
|
|
@ -0,0 +1 @@
|
|||
../../wsj/s5/utils
|
||||
Загрузка…
Ссылка в новой задаче