Difference in pocketsphinx_batch and sphinxtrain

[Ahmed Memon]

Hi,

I've successfully build accoustic model.

SENTENCE ERROR: 16.7% (2/12) WORD ERROR RATE: 5.6% (2/36)

But when I copy hmm, dic & lm.dmp and use with full wav file it decreases the accuracy of wav file with complete voicemail.

word_align produce following:
Insertions: 0 Deletions: 9 Substitutions: 7
TOTAL Words: 42 Correct: 26 Errors: 16
TOTAL Percent correct = 61.90% Error = 38.10% Accuracy = 61.90%
TOTAL Insertions: 0 Deletions: 9 Substitutions: 7

https://www.dropbox.com/s/6edye9qcnpd8ip5/gv2_issue.zip?dl=0 contains all the files required.

Appreciate your help.

[Ahmed Memon]

Update: I've tried adapting the newly created accoustic model, it was displaying FATAL_ERROR: "tmat.c", line 311: Tmat not upper triangular error while I was trying to use run following;

pocketsphinx_continuous -hmm gv2-adapt -lm gv2.lm.dmp -dict gv2.dic -infile wav/4.wav

So I just used transition_matrices from model/en-us/en-us and placed in my gv2-adapt directory, and result were same as above.

[Nickolay V. Shmyrev]

All those errors are due to insufficient data for training. Data requirements are listed in tutorial.