Unsupervised Training with HMMIRest

Acoustic Model Discussions

Flat

User: khemiri
Date: 8/23/2011 7:53 am

Views: 5548
Rating: 4

Hello, I want to perform discriminative unsupervised training usinh HMMIRest, but I don't know how to begin. I read in HTKBOOK that I need to create a denominator word lattices which represent the set of most likeliy word sequences computed by recognition models and a Numerator word lattices which represent likelihood of data given correct transcription. But since I am performing unsupervised training I don't have the correct transcriptions. Any ideas? Thank you very much in advance.

--- (Edited on 8/23/2011 7:53 am [GMT-0500] by ) ---

Re: Unsupervised Training with HMMIRest

User: TonyR
Date: 8/23/2011 9:46 am

Views: 166
Rating: 3

Why do you ask? My feeling is that it if it was a research topic you should be asking different questions and if you've been tasked to do this then you probably need help from someone with more experiance.

--- (Edited on 23-August-2011 3:46 pm [GMT+0100] by TonyR) ---

Re: Unsupervised Training with HMMIRest

User: khemiri
Date: 8/29/2011 7:10 am

Views: 77
Rating: 2

Hi,

I am just curious about that, before I was using HEREST to perform HMM unsupervised training and it works pretty well and now I want just to test the disciminative training and see if there is a big difference.

--- (Edited on 8/29/2011 7:11 am [GMT-0500] by khemiri) ---

Re: Unsupervised Training with HMMIRest

User: TonyR
Date: 8/29/2011 10:06 am

Views: 2685
Rating: 3

Well, you've acomplished one big hurdle, that of unsupervised ML training - I'd be interested to how you did that.

The next thing to do is to follow the HTK book and see if you can get decent supervised dicriminative models (i.e. in line with the improvements others report).

Then put the two halves together and I hope it all works for you. You may not get all the gains as (persumably) you've got much looser denominator lattices. If you are looking for people who've done this before then the HTK users list is probably best.

Good luck,

Tony

--- (Edited on 29-August-2011 4:06 pm [GMT+0100] by TonyR) ---

Previous • Next •


Username	Password