hoes htk2s3conv really work?

[Rafael Oliveira]

Hi,

It's been a while since I've been trying to use the tool "htk2s3conv" aiming
to convert my htk models to sphinx format, but unfortunately until the present
moment my efforts were in vain.

Nowadays, I'm able to use the tool to convert my models well (I already know
how to prepare my htk models to be converted by the tool), but none of this
models really work.

I know that make the front-end configurations compatible is necessary and I've
been working on it. My last effort attempting to do it was test the front-end
configurations bellow (that I found here http://nsh.nexiwave.com/2009/09
/using-htk-models-in-sphinx4.html), but without success =/

HTK:

SOURCEFORMAT = WAV
TARGETKIND = MFCC_D_A_Z_0
TARGETRATE = 100000.0
WINDOWSIZE = 250000.0
USEHAMMING = T
PREEMCOEF = 0.97
NUMCHANS = 26
CEPLIFTER = 22
NUMCEPS = 12
ENORMALISE = T
ZMEANSOURCE = T
USEPOWER = T

Sphinx

-round_filters no
-unit_area no
-remove_dc yes
-transform htk
-lifter 22
-nfilt 26
-lowerf 1
-upperf 8000

So, what I would like to know is: Does anyone here has managed to use
htk2s3conv? Does it really work?

[Nickolay V. Shmyrev]

TARGETKIND = MFCC_D_A_Z_0

This creates wrong order of the features I suppose. Default 1s_c_d_dd assumes
that we have 13 cepstrum values first, then 13 deltas and then 13 delta-
deltas. MFCC_D_A_Z_0 should be something different.

Does anyone here has managed to use htk2s3conv?

Yes

Does it really work?

Yes

[Rafael Oliveira]

Thanks nshmyrev,

This creates wrong order of the features I suppose. Default 1s_c_d_dd
assumes that we have 13 cepstrum values first, then 13 deltas and then 13
delta-deltas. MFCC_D_A_Z_0 should be something different.

I tested your supposition and the files created with both MFCC_D_A_Z_0 and
MFCC_0_D_A_Z were exactly the same (http://www.laps.ufpa.br/pedrobatista/mfc/
). I must be something else.

Does anyone here has managed to use htk2s3conv?
yes

Were this results published? Or is there any documentation I could use to
reproduce these experiments?

[Nickolay V. Shmyrev]

If you have some issues you can try with simple example first.

You can try to decode mfcc files created by HCopy to make sure model is
properly converted.

You can always share the example which can help to reproduce the problem.

[Rafael Oliveira]

You can try to decode mfcc files created by HCopy to make sure model is
properly converted.

How? Converting the mfcc files created by HCopy to the Sphinx format? Or is
there a way I can directly use it on Sphinx?

I tried once a time to convert the mfcc files created by HCopy to the sphinx
format, but only the 13 cepstrum, having in mind that sphinx' s decoders
calculate the 13 deltas and 13 delta-deltas on the fly. In the end I coulnd' t
realize a configuration set to do the decoders calculate the same 13 deltas
and 13 delta-deltas HCopy calculated.

Now I have one more question: Is it possible to feed the sphinx3 with a full
mfcc file ( with the 13 cepstrum 13 deltas and 13 delta-deltas) instead let It
calculate the 13 deltas and 13 delta-deltas on the fly?

[Nickolay V. Shmyrev]

Converting the mfcc files created by HCopy to the Sphinx format? Or is there
a way I can directly use it on Sphinx?

You can write code to do both. It should be simple

Now I have one more question: Is it possible to feed the sphinx3 with a full
mfcc file ( with the 13 cepstrum 13 deltas and 13 delta-deltas) instead let It
calculate the 13 deltas and 13 delta-deltas on the fly?

Yes, why not. Use options "-feat 1s_c -ceplen 39"