Added comment to run.sh; minor fix to decode_sgmmb.sh (added gselect)

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@64 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

egs/rm/s1/run.sh

@@ -97,7 +97,16 @@ done
 
 steps/train_ubma.sh
 
-# train and test unadapted system
+# note: if the SGMM decoding is too slow, aside from playing
+# with decoder beams and max-leaves, you can set e.g.
+# --full-gmm-nbest=5 to the sgmm-gselect program (default is 15, 
+# so max possible speedup with this setting is 3x).  For best WER,
+# this should have the
+# same value in training and test ("matched training"), but
+# you can get the speed improvements by just doing it in test.
+# You can take this all the way down to 1 for fastest speed, although
+# this will degrade results.
+
 (steps/train_sgmma.sh; steps/decode_sgmma.sh)&
 
 # train and test system with speaker vectors.

egs/rm/s1/steps/decode_sgmmb.sh

@@ -54,7 +54,7 @@ for test in mar87 oct87 feb89 oct89 feb91 sep92; do
        ark:$dir/test_${test}.vecs ) 2>$dir/vecs_${test}.log
   
 
-  sgmm-decode-faster $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $model $graphdir/HCLG.fst "$feats" ark,t:$dir/test_${test}.tra ark,t:$dir/test_${test}.ali  2> $dir/decode_${test}.log
+  sgmm-decode-faster "$gselect_opt" $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $model $graphdir/HCLG.fst "$feats" ark,t:$dir/test_${test}.tra ark,t:$dir/test_${test}.ali  2> $dir/decode_${test}.log
 
  # the ,p option lets it score partial output without dying..
   scripts/sym2int.pl --ignore-first-field data/words.txt data_prep/test_${test}_trans.txt | \

egs/wsj/s1/run.sh

@@ -116,7 +116,6 @@ done
 cat data/train.txt | scripts/sym2int.pl --ignore-first-field data/words.txt  > data/train.tra
 
 
-
 # Get the right paths on our system by sourcing the following shell file
 # (edit it if it's not right for your setup). 
 . path.sh