зеркало из https://github.com/mozilla/kaldi.git
Script changes (esp. RE SGMMs).
git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@10 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8
This commit is contained in:
Dan Povey
Родитель
e92655c7be
Коммит
4330846ff0
|
|
@ -38,7 +38,6 @@ cd data_prep
|
|||
./run.sh /path/to/RM
|
||||
cd ..
|
||||
|
||||
|
||||
mkdir -p data
|
||||
( cd data; cp ../data_prep/{train,test*}.{spk2utt,utt2spk} . ; cp ../data_prep/spk2gender.map . )
|
||||
|
||||
|
|
@ -50,14 +49,14 @@ steps/prepare_graphs.sh
|
|||
# data to (e.g. make it a link to a file on some reasonably large file system).
|
||||
# If it doesn't exist, the scripts below will make the directory "exp".
|
||||
|
||||
# tempdir should be set to some place to put training mfcc's
|
||||
# mfcc should be set to some place to put training mfcc's
|
||||
# where you have space.
|
||||
#e.g.: tempdir=/mnt/matylda6/jhu09/qpovey/kaldi_rm_mfccb
|
||||
#e.g.: mfccdir=/mnt/matylda6/jhu09/qpovey/kaldi_rm_mfccb
|
||||
mfccdir=/path/to/mfccdir
|
||||
steps/make_mfcc_train.sh $mfccdir
|
||||
steps/make_mfcc_test.sh $mfccdir
|
||||
|
||||
steps/train_mono.sh
|
||||
steps/train_mono.sh
|
||||
steps/decode_mono.sh &
|
||||
steps/train_tri1.sh
|
||||
steps/decode_tri1.sh &
|
||||
|
|
|
|||
|
|
@ -34,8 +34,6 @@ scripts/mkgraph.sh $tree $model $graphdir
|
|||
for test in mar87 oct87 feb89 oct89 feb91 sep92; do
|
||||
(
|
||||
# Comment the two lines below to make this per-utterance.
|
||||
# This would only work if $srcdir was also per-utterance [otherwise
|
||||
# you'd have to mess with the script a bit].
|
||||
spk2utt_opt=--spk2utt=ark:data/test_${test}.spk2utt
|
||||
utt2spk_opt=--utt2spk=ark:data/test_${test}.utt2spk
|
||||
|
||||
|
|
|
|||
|
|
@ -38,8 +38,6 @@ scripts/mkgraph.sh $tree $model $graphdir
|
|||
for test in mar87 oct87 feb89 oct89 feb91 sep92; do
|
||||
(
|
||||
# Comment the two lines below to make this per-utterance.
|
||||
# This would only work if $srcdir was also per-utterance [otherwise
|
||||
# you'd have to mess with the script a bit].
|
||||
spk2utt_opt=--spk2utt=ark:data/test_${test}.spk2utt
|
||||
utt2spk_opt=--utt2spk=ark:data/test_${test}.utt2spk
|
||||
|
||||
|
|
|
|||
|
|
@ -1,48 +0,0 @@
|
|||
#!/bin/bash
|
||||
# Copyright 2010-2011 Microsoft Corporation
|
||||
|
||||
# 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.
|
||||
|
||||
|
||||
# Initialize SGMM from a trained HMM/GMM system.
|
||||
|
||||
if [ -f path.sh ]; then . path.sh; fi
|
||||
|
||||
dir=exp/sgmm/init
|
||||
mkdir -p $dir
|
||||
srcdir=exp/tri1
|
||||
model=exp/sgmm/0.mdl
|
||||
|
||||
init-ubm --intermediate-numcomps=2000 --ubm-numcomps=400 --verbose=2 \
|
||||
--fullcov-ubm=true $srcdir/final.mdl $srcdir/final.occs \
|
||||
$dir/ubm0 2> $dir/cluster.log
|
||||
|
||||
|
||||
subset[0]=1000
|
||||
subset[1]=1500
|
||||
subset[2]=2000
|
||||
subset[3]=2500
|
||||
|
||||
for x in 0 1 2 3; do
|
||||
echo "Pass $x"
|
||||
feats="ark:scripts/subset_scp.pl ${subset[$x]} data/train.scp | add-deltas --print-args=false scp:- ark:- |"
|
||||
fgmm-global-acc-stats --diag-gmm-nbest=15 --binary=false --verbose=2 $dir/ubm$x "$feats" $dir/$x.acc \
|
||||
2> $dir/acc.$x.log || exit 1;
|
||||
fgmm-global-est --verbose=2 $dir/ubm$x $dir/$x.acc \
|
||||
$dir/ubm$[$x+1] 2> $dir/update.$x.log || exit 1;
|
||||
rm $dir/$x.acc
|
||||
done
|
||||
|
||||
sgmm-init $srcdir/final.mdl $dir/ubm4 $model 2> $dir/sgmm_init.log
|
||||
|
||||
|
|
@ -28,9 +28,11 @@ numiters=25 # Total number of iterations
|
|||
|
||||
realign_iters="5 10 15";
|
||||
silphonelist=`cat data/silphones.csl`
|
||||
numsubstates=1500 # Initial #-substates.
|
||||
totsubstates=5000 # Target #-substates.
|
||||
numleaves=2500
|
||||
numsubstates=2500 # Initial #-substates.
|
||||
totsubstates=7500 # Target #-substates.
|
||||
maxiterinc=15 # Last iter to increase #substates on.
|
||||
|
||||
incsubstates=$[($totsubstates-$numsubstates)/$maxiterinc] # per-iter increment for #substates
|
||||
gselect_opt="--gselect=ark:gunzip -c $dir/gselect.gz|"
|
||||
randprune=0.1
|
||||
|
|
@ -38,12 +40,12 @@ mkdir -p $dir
|
|||
|
||||
feats="ark:add-deltas --print-args=false scp:data/train.scp ark:- |"
|
||||
|
||||
cp $srcdir/tree $dir
|
||||
cp $srcdir/topo $dir
|
||||
|
||||
if [ ! -f $ubm ]; then
|
||||
echo "No UBM in $ubm"
|
||||
echo "No UBM in $ubm";
|
||||
exit 1
|
||||
fi
|
||||
sgmm-init $srcdir/final.mdl $ubm $dir/0.mdl 2> $dir/sgmm_init.log
|
||||
|
||||
echo "aligning all training data"
|
||||
if [ ! -f $dir/0.ali ]; then
|
||||
|
|
@ -51,6 +53,33 @@ if [ ! -f $dir/0.ali ]; then
|
|||
"$feats" ark,t:$dir/0.ali 2> $dir/align.0.log || exit 1;
|
||||
fi
|
||||
|
||||
# We rebuild the tree because we want a larger #states than for a normal
|
||||
# GMM system (the optimum #states for SGMMs tends to be a bit higher).
|
||||
|
||||
if [ ! -f $dir/treeacc ]; then
|
||||
acc-tree-stats --ci-phones=$silphonelist $srcmodel "$feats" ark:$dir/0.ali \
|
||||
$dir/treeacc 2> $dir/acc.tree.log || exit 1;
|
||||
fi
|
||||
|
||||
cat data/phones.txt | awk '{print $NF}' | grep -v -w 0 > $dir/phones.list
|
||||
cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/questions.log || exit 1;
|
||||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
$dir/questions.qst $dir/topo $dir/tree 2> $dir/train_tree.log || exit 1;
|
||||
|
||||
# the sgmm-init program accepts a GMM, so we just create a temporary GMM "0.gmm"
|
||||
|
||||
gmm-init-model --write-occs=$dir/0.occs \
|
||||
$dir/tree $dir/treeacc $dir/topo $dir/0.gmm 2> $dir/init_gmm.log || exit 1;
|
||||
|
||||
sgmm-init $dir/0.gmm $ubm $dir/0.mdl 2> $dir/init_sgmm.log || exit 1;
|
||||
|
||||
rm $dir/0.gmm
|
||||
|
||||
if [ ! -f $dir/gselect.gz ]; then
|
||||
sgmm-gselect $dir/0.mdl "$feats" ark,t:- 2>$dir/gselect.log | gzip -c > $dir/gselect.gz || exit 1;
|
||||
fi
|
||||
|
|
@ -76,13 +105,12 @@ while [ $iter -lt $numiters ]; do
|
|||
sgmm-acc-stats-ali --update-flags=$flags "$gselect_opt" --rand-prune=$randprune --binary=false $dir/$iter.mdl "$feats" ark:$dir/cur.ali $dir/$iter.acc 2> $dir/acc.$iter.log || exit 1;
|
||||
sgmm-est --update-flags=$flags --split-substates=$numsubstates --write-occs=$dir/$[$iter+1].occs $dir/$iter.mdl $dir/$iter.acc $dir/$[$iter+1].mdl 2> $dir/update.$iter.log || exit 1;
|
||||
fi
|
||||
# TEMP: will restore these statements later.
|
||||
# rm $dir/$iter.mdl $dir/$iter.acc
|
||||
# rm $dir/$iter.occs
|
||||
if [ $iter -lt $maxiterinc ]; then
|
||||
numsubstates=$[$numsubstates+$incsubstates]
|
||||
fi
|
||||
iter=$[$iter+1];
|
||||
rm $dir/$iter.mdl $dir/$iter.acc
|
||||
rm $dir/$iter.occs
|
||||
if [ $iter -lt $maxiterinc ]; then
|
||||
numsubstates=$[$numsubstates+$incsubstates]
|
||||
fi
|
||||
iter=$[$iter+1];
|
||||
done
|
||||
|
||||
( cd $dir; rm final.mdl final.occs 2>/dev/null; ln -s $iter.mdl final.mdl; ln -s $iter.occs final.occs )
|
||||
|
|
|
|||
|
|
@ -33,8 +33,9 @@ ubm=exp/ubma/4.ubm
|
|||
realign_iters="5 10 15";
|
||||
spkvec_iters="5 8 12 17 22"
|
||||
silphonelist=`cat data/silphones.csl`
|
||||
numsubstates=1500 # Initial #-substates.
|
||||
totsubstates=5000 # Target #-substates.
|
||||
numleaves=2500
|
||||
numsubstates=2500 # Initial #-substates.
|
||||
totsubstates=7500 # Target #-substates.
|
||||
maxiterinc=15 # Last iter to increase #substates on.
|
||||
incsubstates=$[($totsubstates-$numsubstates)/$maxiterinc] # per-iter increment for #substates
|
||||
gselect_opt="--gselect=ark:gunzip -c $dir/gselect.gz|"
|
||||
|
|
@ -52,9 +53,7 @@ if [ ! -f $ubm ]; then
|
|||
echo "No UBM in $ubm"
|
||||
fi
|
||||
|
||||
sgmm-init --spk-space-dim=39 $srcdir/final.mdl $ubm $dir/0.mdl 2> $dir/sgmm_init.log || exit 1;
|
||||
|
||||
cp $srcdir/tree $dir
|
||||
cp $srcdir/topo $dir
|
||||
|
||||
echo "aligning all training data"
|
||||
if [ ! -f $dir/0.ali ]; then
|
||||
|
|
@ -62,8 +61,33 @@ if [ ! -f $dir/0.ali ]; then
|
|||
"$feats" ark,t:$dir/0.ali 2> $dir/align.0.log || exit 1;
|
||||
fi
|
||||
|
||||
# We rebuild the tree because we want a larger #states than for a normal
|
||||
# GMM system (the optimum #states for SGMMs tends to be a bit higher).
|
||||
|
||||
if [ ! -f $dir/treeacc ]; then
|
||||
acc-tree-stats --ci-phones=$silphonelist $srcmodel "$feats" ark:$dir/0.ali \
|
||||
$dir/treeacc 2> $dir/acc.tree.log || exit 1;
|
||||
fi
|
||||
|
||||
cat data/phones.txt | awk '{print $NF}' | grep -v -w 0 > $dir/phones.list
|
||||
cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/questions.log || exit 1;
|
||||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
$dir/questions.qst $dir/topo $dir/tree 2> $dir/train_tree.log || exit 1;
|
||||
|
||||
# the sgmm-init program accepts a GMM, so we just create a temporary GMM "0.gmm"
|
||||
|
||||
gmm-init-model --write-occs=$dir/0.occs \
|
||||
$dir/tree $dir/treeacc $dir/topo $dir/0.gmm 2> $dir/init_gmm.log || exit 1;
|
||||
|
||||
sgmm-init --spk-space-dim=39 $dir/0.gmm $ubm $dir/0.mdl 2> $dir/init_sgmm.log || exit 1;
|
||||
|
||||
if [ ! -f $dir/0.mdl ]; then
|
||||
echo "you must run init_sgmm.sh before train_sgmm1.sh"
|
||||
echo "you must run init_sgmm.sh before train_sgmmb.sh"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
|
|
|
|||
|
|
@ -58,7 +58,7 @@ scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.tx
|
|||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
# Have to make silence root not-shared because we will not split it.
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
|
|
|
|||
|
|
@ -55,7 +55,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -78,7 +78,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -77,7 +77,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -56,7 +56,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -63,7 +63,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -65,7 +65,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -82,7 +82,7 @@ done
|
|||
|
||||
|
||||
# just a single element. :-separated integer list of context-independent
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
# script below tells it not to cluster, but here we avoid accumulating
|
||||
# CD-stats for silence.
|
||||
|
||||
|
|
@ -106,7 +106,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -65,7 +65,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -69,7 +69,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -66,7 +66,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
@ -182,7 +182,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -92,7 +92,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
|
|
|
|||
|
|
@ -68,7 +68,7 @@ cluster-phones $dir/treeacc $dir/phones.list $dir/questions.txt 2> $dir/question
|
|||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
compile-questions $dir/topo $dir/questions.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.pl --separate data/phones.txt `cat data/silphones.csl` shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
scripts/make_roots.pl --separate data/phones.txt $silphonelist shared split > $dir/roots.txt 2>$dir/roots.log || exit 1;
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
|
|
|
|||
|
|
@ -37,9 +37,9 @@ subset[3]=2500
|
|||
for x in 0 1 2 3; do
|
||||
echo "Pass $x"
|
||||
feats="ark:scripts/subset_scp.pl ${subset[$x]} data/train.scp | add-deltas --print-args=false scp:- ark:- |"
|
||||
fgmm-acc-stats --diag-gmm-nbest=15 --binary=false --verbose=2 $dir/$x.ubm "$feats" $dir/$x.acc \
|
||||
fgmm-global-acc-stats --diag-gmm-nbest=15 --binary=false --verbose=2 $dir/$x.ubm "$feats" $dir/$x.acc \
|
||||
2> $dir/acc.$x.log || exit 1;
|
||||
fgmm-est --verbose=2 $dir/$x.ubm $dir/$x.acc \
|
||||
fgmm-global-est --verbose=2 $dir/$x.ubm $dir/$x.acc \
|
||||
$dir/$[$x+1].ubm 2> $dir/update.$x.log || exit 1;
|
||||
rm $dir/$x.acc $dir/$x.ubm
|
||||
done
|
||||
|
|
|
|||
|
|
@ -84,12 +84,14 @@ cat data/lexicon.txt | awk '{print $1}' | sort | uniq | \
|
|||
|
||||
cd data_prep
|
||||
|
||||
#TODO: remove following system-specific comments.
|
||||
#On BUT system, do:
|
||||
./run.sh /mnt/matylda2/data/WSJ?/??-{?,??}.?
|
||||
|
||||
# On Geoff Hinton's system we can do:
|
||||
# ./run.sh /ais/gobi2/speech/WSJ/*/??-{?,??}.?
|
||||
# On BUT system, do:
|
||||
# The following command needs a list of directory names from
|
||||
# the LDC's WSJ disks. These will end in e.g. 11-1.1.
|
||||
# examples:
|
||||
# /ais/gobi2/speech/WSJ/*/??-{?,??}.?
|
||||
# /mnt/matylda2/data/WSJ?/??-{?,??}.?
|
||||
./run.sh [list-of-directory-names]
|
||||
|
||||
|
||||
cd ..
|
||||
|
|
|
|||
1
src/TODO
1
src/TODO
|
|
@ -6,6 +6,7 @@
|
|||
|
||||
TODO items (mainly for Dan to do):
|
||||
|
||||
Add separate min-count at root of tree for regression-tree fMLLR/MLLR
|
||||
Add fMLLR scripts for SGMM, and also use deeper trees.
|
||||
Document configure script.
|
||||
Remove cpplint.py from distribution.
|
||||
|
|
|
|||
|
|
@ -20,7 +20,7 @@
|
|||
namespace kaldi {
|
||||
|
||||
/**
|
||||
\page model Acoustic modeling code
|
||||
\page model Acoustic modeling code
|
||||
|
||||
\section model_intro Introduction
|
||||
|
||||
|
|
@ -29,7 +29,7 @@ namespace kaldi {
|
|||
models (i.e. diagonal GMMs) and Subspace Gaussian Mixture Models (SGMMs), but
|
||||
also to be easily extensible to new kinds of model. In a previous iteration of
|
||||
designing this software, we used a virtual base class that both the GMM and
|
||||
SGMM classes inherited from, and wrote command-line tools to handle both types of
|
||||
SGMM classes inherited from, and wrote command-line tools that handled both types of
|
||||
model. Our experience was that a base class is not as useful as one might
|
||||
think, because there are too many differences in the models (e.g. they support
|
||||
different types of adaptation), and we were forced to constantly expand the
|
||||
|
|
@ -100,12 +100,22 @@ namespace kaldi {
|
|||
for such things as model estimation (e.g. see MlEstimateAmDiagGmm), or transform
|
||||
estimation (there are various pieces of code that do this; see \ref transforms.
|
||||
|
||||
\subsection model_full_gmm Full-covariance GMMs
|
||||
|
||||
We have a class \ref FullGmm for full-covariance GMMs, which has similar functionality
|
||||
to the \ref DiagGmm class but with full covariances. This is mainly of use for training
|
||||
full-covariance Universal Background Models (UBMs) in the SGMM recipe (see below).
|
||||
The only command-line tools available for full GMMs are used to train global mixture models
|
||||
(i.e. UBMs); we have not
|
||||
implemented a full covariance version of the AmDiagGmm class or the corresponding command
|
||||
line tools, although doing so would be fairly easy.
|
||||
|
||||
\section model_sgmm Subspace Gaussian Mixture Models (SGMMs)
|
||||
|
||||
Subspace Gaussian Mixture Models (SGMMs) are implemented by class
|
||||
AmSgmm. This class essentially implements the approach described in
|
||||
``The Subspace Gaussian Mixture Model – a Structured Model for Speech
|
||||
Recognition'', by D. Povey, Lukas Burget et. al Computer Speech and Language,
|
||||
``The Subspace Gaussian Mixture Model -- a Structured Model for Speech
|
||||
Recognition'', by D. Povey, Lukas Burget et. al, Computer Speech and Language,
|
||||
2011.
|
||||
The class AmSgmm represents a whole collection of pdf's; there
|
||||
is no class that represents a single pdf of the SGMM (as there is for
|
||||
|
|
@ -118,3 +128,4 @@ namespace kaldi {
|
|||
|
||||
|
||||
*/
|
||||
}
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче