[hts-users:03264] Re: Regarding to the adaptation part of the demo

[li jay <lij.acd@xxxxxxxxx>]

Hi,

Thank you for your advise. I've tried the SAT training and adaptation, it did perform better than the one of SI model.

Jay

那兴宇 <nxy-yzqs@xxxxxxx> 於 2012年4月20日上午10:33 ?道：

Hi,

SI training provides the basic model and regression-class tree for SAT training.

But you do not have to do that again, just use the SI model (including unseen models) and the regression-class tree you have already got.

--

Xingyu Na (那兴宇)

Beijing Institute of Technology

naxy(at)bit.edu.cn

asr.naxingyu(at)gmail.com

naxingyu at {facebook, twitter, linkedin}

At 2012-04-20 00:03:00,"li jay" <lij.acd@xxxxxxxxx> wrote:

Hi,

Thank you for telling me this. I read the paper of Dr. Yamagishi, and I realized that speaker independent training  was a conventional way and speaker adaptive training (SAT) was a new one. I think that's why my generated adapted voice was not so good.

But there is something still confusing me: 

My training process used to train the average model (I call it the basic training step) was:

# preparing environments

# HCompV (computing variance floors)

# HInit & HRest (initialization & reestimation)

# HHEd (making a monophone mmf)

# HERest (embedded reestimation (monophone))

# HHEd (copying monophone mmf to fullcontext one)

# HERest (embedded reestimation (fullcontext))

# HHEd (tree-based context clustering)

# HERest (embedded reestimation (clustered))

# HHEd (untying the parameter sharing structure)

# HERest (embedded reestimation (untied))

# HHEd (tree-based context clustering)

# HERest (embedded reestimation (re-clustered))

# HFst (forced alignment using WFST for no-silent GV)

, which are in the order of what the script do, and they work. After running script above, I performed 1~5 to generate the SI model adapted voice. 

As to generating SAT model adapted voice, should I still run the script above (the basic training step), then use the script below? 

1 # HHEd (building regression-class trees for adaptation) and 6~9 and 10~13

Why I am confusing is because based on the paper of Dr. Yamagishi, it seems like that the training part of average voice model is different from the conventional one. I don't know if the script (the basic training step) is suitable or not.

Jay

<nxy-yzqs@xxxxxxx> 於 2012年4月19日上午12:55 ?道：

Hi,

1~5 is adaptation based on SI model.
6~9 is speaker adaptive training for average voice model
10~13 is adaptation based on average voice model
For reference, please read Dr. Yamagishi's papers on the publication list of HTS website.


在 2012-04-18 20:24:05，"li jay" <lij.acd@xxxxxxxxx> 写道：

Hi,

I want to ask something regarding to adaptation part of HTS-demo_CMU-ARCTIC-ADAPT demo Training.pl script.

I used sentences from several speakers to train a average model, and then used the following parts (1~5) of codes to adapt to specific speaker and generate voices:

1 # HHEd (building regression-class trees for adaptation)

2 # HERest (speaker adaptation (speaker independent))

3 # HERest (speaker adaptation (SI+MLLR+MAP))

4 # HMGenS (generating speech parameter sequences (speaker adapted))

5 # SPTK (synthesizing waveforms (speaker adapted))

The generated adapted voice was ok, but not so good. I want to ask what the following parts (6~9 and 10~13) are for? 

6 # HERest (Speaker adaptive training (SAT))

7 # HHEd (making unseen models (SAT))

8 # HMGenS (generating speech parameter sequences (SAT))

9 # SPTK (synthesizing waveforms (SAT))

and 

10 # HERest (speaker adaptation (SAT))

11 # HERest (speaker adaptation (SAT+MLLR+MAP))

12 # HMGenS (generate speech parameter sequences (SAT+adaptation))

13 # SPTK (synthesizing waveforms (SAT+adaptation))

They all seem like adaptation and voice generation. What is the difference between them(1~5, 6~9, and 10~13)?

Jay