[hts-users:00024] Re: MSDINFO Parameters...

[Heiga ZEN <zen@xxxxxxxxxxxxxxxx>]

Hi Elliot,

On Fri, 2 Apr 2004 10:18:08 -0500
"Elliot Moore" <em80@xxxxxxxxxxxxxxx> wrote:

> I am bulding a monophone recognizer using mel-cepstral coefficients of order 24.
> I have read the paper on MSD-HMM's and I want to make sure 
> I understand it's correct implementation into the modified HMM prototype definition for HTK.  
> Is there any documentation that I have missed that explains the <MSDINFO> line?

Unfortunately, there is no documentation about the modified prototype definition.

> From the  demo I extracted this line from the prototype:
> ~o <VecSize> 78 <USER> <MSDINFO> 4 0 1 1 1 <StreamInfo> 4 75 1 1 1
>
> 1. From looking at the code I can tell that the 4 represents the number of streams 
>    where the first stream is the spectrum and the last three belong to logf0 parameters.

Yes, that's right.
  
>    However, what do the 0 and 1's mean?  
>    I assumed that 0 meant 0-dimensions or discrete symbol.  
>    But it is in the first stream position for the spectrum which is a continuous distribution.

No, "0" doesn't mean 0-dimension or discrete.
Symbol "0" means that this stream is not modeled by multi-space probability distribution (MSD) 
and "1" means that this stream is modeled by MSD.

If you don't use MSD, <MSDINFO> does not required.

> For my work, I am using a 5 state (3 emitting), 32 mixture, 
> single stream model with mel-cepstral, delta, and delta_delta parameters (75 in all).  
> I used the following line in my prototype definition
> 
> ~o <VecSize> 75 <USER> <MSDINFO> 1 1 <StreamInfo> 1 75
> <BeginHMM>
>   <NumStates> 5
>   <State> 2
>   <Stream> 1
>   <NumMixes> 32
>    <Mixture> 1 0.03125
> etc......

I think there is no problem in this prototype.

> ...And kept getting the following error
> HMM Def Error: GetToken: Symbol expected at line 1/col 2/char 1 in /(dir)/fcjf0-sa1.cmp
>   ERROR [+7050]  HMError:
> HMM Def Error: LoadHMMSet: GetToken failed at line 1/col 0/char -1 /(dir)/fcjf0-sa1.cmp
>   ERROR [+7050]  HMError:
>   ERROR [+2028]  Initialise: LoadHMMSet failed
>  FATAL ERROR - Terminating program /cluster/csip/emoore/htk/bin.linux/HCompV

Did you execute HCompV with proper command line options ?
I think that HCompV confused the training data file (fcjf0-sa1.cmp) with the HMM definition files.
 
Best regards

Heiga Zen

-- 
 ------------------------------------------------
  Heiga Zen     (in Japanese pronunciation)
  Byung-Ha Chun (in Korean pronunciation)
  
  Department of Computer Science and Engineering
  Graduate School of Engineering
  Nagoya Institute of Technology
  Japan

  e-mail: zen@xxxxxxxxxxxxxxxx
     web: http://kt-lab.ics.nitech.ac.jp/~zen
 ------------------------------------------------