[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:02867] Re: Question about duration model structure


Hi,

ZhuFengyun wrote (2011/05/25 6:40):

Dear HTS developers:
As Dr. ZEN said in http://hts.sp.nitech.ac.jp/hts-users/spool/2007/msg00680.html
<http://hts.sp.nitech.ac.jp/hts-users/spool/2007/msg00680.html>
"In the previous versions we used a multi-variate Gaussian PDF to represent
state duration PDFs.
However, since version 2.0.1 we have changed it to multi-stream structure."
I'm confused about why are you making the decision that using multi-stream
structure instead of multi-variate Gaussian PDF?
Can someone make a brief explanation about the advantages of using a
multi-stream structure?

Because implementation of HSMM training and adaptation was easier. With the multi-stream structure, existing functions (SOutP(), ApplyCompFXForm(), ApplyXForm(), etc.) could be reused to implement HSMM training & adaptation.

If we kept using the old multi-variate Gaussian PDF, I needed to implement various additional functions from the scratch for this functionality. The gap between HTK and HTS might increase and maintenance of HTS might be harder.

Regards,

Heiga ZEN (Byung Ha CHUN)

======================================================================
NOTE: The information in this email and any attachments may be confidential
and/or legally privileged. This message may be read, copied and used only by
the intended recipient. If you are not the intended recipient, please
destroy this message, delete any copies held on your system and notify the
sender immediately.

Toshiba Research Europe Limited, registered in England and Wales (2519556).
Registered Office: 208 Cambridge Science Park, Milton Road, Cambridge
CB4 0GZ, England. Web: http://www.toshiba-europe.com/research/crl
======================================================================

______________________________________________________________________
This email has been scanned by the MessageLabs Email Security System.
For more information please visit http://www.messagelabs.com/email ______________________________________________________________________
References
[hts-users:02866] Question about duration model structure, ZhuFengyun