Speech Recognition Engines

Flat
multiple streams
User: Sobh
Date: 1/30/2013 12:06 pm
Views: 4469
Rating: 6

This is copied.

Hi everybody,

I'm using HTK with features extracted from an external model. So far I have been only using a single streams, since I had almost uncorrelated features (and it was working fine).

Anyway, now I'd like to extend my work by inserting features extracted from different channels that are likely to be correlated.

Thus, I wanted to ask if anybody knows if I can use multiple streams (to treat the correlated features) with the features I am providing (which are non-HTK features or user defined features). 

And if that's possible, is there something to be aware of in the different commands I use?

I mean, I basically use the following HTK commands in my script:

HCompV

HERest

HVite

HResults

I would expect that the only modifications that I have to introduce in the scripts that were working for the 'single stream' case are only in the prototypes definition of the HMMs (since I have to define number of streams, streams' features number and streams' weights) and nothing else.

Or do you think there are other parameter of the aforementioned commands I should edit?

 

 

--- (Edited on 1/30/2013 12:06 pm [GMT-0600] by Visitor) ---

Re: multiple streams
User: TonyR
Date: 1/30/2013 12:38 pm
Views: 2125
Rating: 2

Streams in HTK are independent so probably won't get you want you want.  Have you considered using HLDA to decorrelate your data?   Or, even simpler, you could scale your data by the variances and then use PCA to get something that HTK can model with diagonal covariances.

 

Tony

-- 

Dr Tony Robinson
CEO Cantab Research Ltd

--- (Edited on 30-January-2013 6:38 pm [GMT+0000] by TonyR) ---

PreviousNext