[hts-users:01527] Re: Can HMGenS do state alignment from extracted feature data?

[John Dines <dines@xxxxxxxx>]

Hi
>
> Thanks very much for the reply. I've checked the ML archive and tried
> with Simon's method. However, there is still 2 points I doubts.
>
> 1. The equivalence of the re-estimated transition matrices comparing
> with duration models, while it performs in the viterbi state alignment.
>
They are not equivalent, but generally, the likelihoods of the
observation vectors dominate the computation of the state occupancy
probabilities in the Baum-Welch algorithm. So your alignment should
still be reliable, even though it is performed without HSMM duration
models.
>
> 2. How to only update transition matrices with HTS modified HERest?
> Seems it needs duration models to compute with composite HMM models.
> And the updated transition matrices are not updated from training data
> instead of a dummy one. Is there any trigger to turn off the duration
> model computation and back to normal HMM update computation in HERest
> with HTS-2.1 or any other old version?
>
1) First you need to decouple the 'cmp' models and the 'dur' models eg.
run HHEd with the following edit commands (as also described by Simon)
on the MMF containing the 'cmp' HMM statistics:

TR 2

UT {*.transP}
TI T_aa {(*-aa+*,aa,*-aa,aa+*).transP}
TI T_ae {(*-ae+*,ae,*-ae,ae+*).transP}
...
TI T_zh {(*-zh+*,zh,*-zh,zh+*).transP}
AT 2 2 0.9 {*.transP}
AT 2 3 0.1 {*.transP}
AT 3 3 0.9 {*.transP}
AT 3 4 0.1 {*.transP}
AT 4 4 0.9 {*.transP}
AT 4 5 0.1 {*.transP}
AT 5 5 0.9 {*.transP}
AT 5 6 0.1 {*.transP}
AT 6 6 0.9 {*.transP}
AT 6 7 0.1 {*.transP}

2) Retrain the transition matrices in the 'cmp' MMF using HERest and the
'-u t' switch which should only update the transition matrices.

3) Run HVite with '-a' switch to perform alignment.

John
> Thanks
>
> On Fri, Jul 4, 2008 at 11:55 PM, Heiga ZEN (Byung Ha CHUN)
> <zen@xxxxxxxxxxxxxxx <mailto:zen@xxxxxxxxxxxxxxx>> wrote:
>
>     Hi,
>
>     vancy wrote (2008/07/04 19:33):
>
>
>         I want evaluate the performance and impact of HSMM duration,
>         so I want to do a state alignment from handed label, real
>         pitch and cesptrum data extracted from database, and then
>         generated pitch and cepstral parameter with this state
>         alignment. Could this be done by HMGenS or with any other helps?
>
>
>     Currently HTS doesn't include any tools to perform Viterbi
>     alignment with HSMMs. Internally we have a WFST-based decoder for
>     HSMMs. But it's not open. Simon King has shown an approximate
>     solution for this problem. Please refer to hts-users:01140-01143
>     for details. You can find them in hts-users ML archive
>     (http://hts.sp.nitech.ac.jp/hts-users/spool/2008/threads.html).
>
>     Once state-level alignment is obtained, you can generate cepstrum
>     and F0 parameters with this state alignment using HMGenS with -s
>     option.
>
>     Best regards,
>
>     Heiga ZEN (Byung Ha CHUN)
>