[SPHINX3] : bw & (c)mllr

[svanni]

Hello,

I'm trying to evaluate a two pass system with unsupervised adaptation of means
with MLLR first, and means and variances with CMLLR ( inspired by the python
script "mllr.py" and the paper M.J.F. Gales, "Maximum likehood Linear
Transformations for HMM-Based Speech recognition") later.

I've got a warning from 'bw' when i collect statistics from the hypothesis
done by the first pass , about 1000 s of speech by speaker :

"WARNING: "accum.c", line 627: Over 500 senones never occur in the input data.  This is normal for CD untied training, but could indicate a serious problem otherwise."

I understand that some senones doesn't occur and i hope this a normal issue (
the adaptation is not supervised, i don't know which text is pronounced ).

But i've got to be sure in order to (in)validate my implementation of CMLLR in
Python and since the WER is a little bit higher with CMLLR adapation (MLLR
also).

Thanks by advance for your response,
Stephan

[Nickolay V. Shmyrev]

In trunk it reads:

      E_WARN("Over 500 senones never occur in the input data. "
             "This is normal for context-dependent untied senone training or for adaptation, "
             "but could indicate a serious problem otherwise.\n");

[svanni]

Ok thanks a lot.

I'll be back with my implementation of CMLLR in Python.

Did you interested to get the actual code and try to improve it with me ?

[Nickolay V. Shmyrev]

Did you interested to get the actual code and try to improve it with me ?

We are very interested in contributions!

[svanni]

Ok but it's not really a contribution, because i'm not 100% sure of my code
(the WER is increasing a little bit after adaptation) ...
It's much more a "initialization".
I do some more tests and i come back.
Bye,
Stephan

[Nickolay V. Shmyrev]

Dear Stephan

The faster you show the code the faster we will make it work. :) That's
chicken and egg problem.

[svanni]

Ok thanks Nickolay, the file can be downloaded here :
http://www.mediafire.com/?588en6s1vphs0bd

[asr2010]

Is this feature of two pass decoding has been added to sphinx or not yet?

[asr2010]

By two pass decoding, I mean that while decoding we perform MLLR adapt for
each utterance then we perform decoding.

[Nickolay V. Shmyrev]

Is this feature of two pass decoding has been added to sphinx or not yet? By
two pass decoding, I mean that while decoding we perform MLLR adapt for each
utterance then we perform decoding.

This feature is not supported yet.