[Accessibility] Re: [Ghmm-list] Can you help write a free version of HTK?

[Alexander Schliep]

On Jul 9, 2010, at 11:56 AM, <address@hidden> wrote:
> > http://sourceforge.net/projects/ghmm/
> I have already contacted them a while back and they said that their
> library was not geared towards speech processing and couldn't be  
> used for
> it neither now nor in the short term future IIRC.

That is patently wrong.

What Janne Grunau wrote to you on the GHMM-List 1/19/2009 was:

> From: Janne Grunau <address@hidden>
> Date: 21. Januar 2009 12:33:03 MEZ
> On Monday 19 January 2009 12:01:13 Peter Grasch wrote:
> > just wanted to know if GHMM can be used for that purpose.
>
> I'm not sure if it can be used for speech recognition. The basics HMM
> building blocks are there but I don't know enough about the specific  
> HMM
> usage in speech recognition. Most likely advanced HMMs for speech
> recognition use some "tricks" which aren't implemented in the GHMM.  
> It won't
> be drop in replacement for HTK.
>
> > Looking at your homepage and the sourceforge trove categorization, it
> > seems that the generated models are not meant to be used for speech
> > recognition. Does that pose a problem?
>
> Speech recognition is missing in the categorization since to my  
> knowledge
> nobody uses the GHMM for speech recognition. As the name suggests  
> the GHMM
...

> > Can the GHMM-generated models be
> > used for that? (we use tied-state triphone models)
>
> I don't know what tied-state triphone models are exactly and short  
> internet
> search wasn't conclusive.

...
> specific features are likely to be missing. I can't compare it to  
> HTK since
> even the HTK documentation requires a registration.



AFAIK there was never a follow-up from you. Obviously, the GHMM is not  
a drop-in
replacement for HTK, but there is is nothing which prevents you from  
using it. In fact
there is a very good book by Rabiner about speech recognition laying  
out the details.
Implement an HMM-based speech recognition system with the GHMM would
be absolutely feasible; but as Janne pointed out we do bioinformatics  
not speech
recognition, so it is not done yet. Janne could have been more  
helpful, if at least
the HTK docs would have been out in the open, or if you would have  
explained to him
what the tied-state triphone models. In fact, there are just linear  
HMMs with the emission
distributions tied. We do use tying for discrete emissions, extending  
to continous would
be trivial.

Best,
Alexander



--
NEW Springer Book out: CATBox http://schliep.org/CATBox
Dr. Alexander Schliep, Associate Professor,   address@hidden
Department of Computer Science & BioMaPS Institute for Quantitative  
Biology
Rutgers, The State University of New Jersey, Piscataway, NJ, USA
Tel: (732) 445-2001 ext 1166   Hill Center #264