[hts-users:01220] Re: WER Comparison of HTK and HTS

[Junichi Yamagishi <jyamagis@xxxxxxxxxxxx>]

Hi,

I didn't use any new functions of HTS such as HSMM, CSMAPLR
and so on although I know they could improves WER more.

What I did is to run RMTHTK/RM RECIPE using HTS binary files
instead of HTK binary files (though I slightly changed the
training procedures and configuration variables to reflect our 'HTS  
Recipe').

Thus, in theory, there is no differences between them. What I  
confirmed is
"accuracy" of HTS, in other words, Heiga's magic (bug fix, flooring,  
and so on).

For details, please see RMHTK.

Junichi


On 2008/03/12, at 16:43, Sacha Krstulovic wrote:

>     Hi Junichi;
>
> Junichi Yamagishi wrote:
> > I compared HTS with HTK for speech recognition purpose.
>
> Thanks for announcing this.
>
> Your message may leave the hts-user hungry for
> a bit more information:
>
> What did you use for the decoding: WFSTs or Simon's
> HVite-based recipe? (Or did you not use HSMMs at all?)
>
> Did you use a language model? If yes, what kind of LM?
>
> What kind of questions did you use for the state clustering:
> quinphones only, or quinphones+phonological features, or
> a full set of TTS-like context features, or something else?
>
> Would you please remind us what are the other relevant
> system differences between HTK and HTS, besides
> the MDL criterion and possibly the HSMMs? E.g.,
> did you use semi-tied covariances?
>
> Is it planned to release MPE as part of HTS-2.1?
> Would MPE then apply to the HSMMs as well?
>
>
> Thanks and best regards;
>                                -*- Sacha K. -*-
> --
> Dr.Sacha Krstulovic - Research Engineer
> Toshiba Research Europe Limited
> Cambridge Research Laboratory
> Speech Technology Group
> 208 Science Park, Milton Road
> Cambridge CB4 0GZ - United Kingdom
> Tel:    +44 1223 436 978
> Fax:    +44 1223 436 909
> E-mail: sacha@xxxxxxxxxxxxxxxxx
>
>
>
> ______________________________________________________________________
> This email has been scanned by the MessageLabs Email Security System.
> For more information please visit http://www.messagelabs.com/email  
> ______________________________________________________________________ 
> <sacha.vcf>