Click here to register.

General Discussion

Flat
Training speaker dependent models from scratch
User: nestea247
Date: 3/1/2008 5:53 pm
Views: 2817
Rating: 14

I have some english audio lectures from my professor at school I want to use Julius as a continuous dictation ASR to create some rough transcripts for. The lectures have technical terms that a general english language model/dictionary will not suffice.

How should I start (from scratch)? What's existing in the public domain that I can use?

How do I create a language model? How do I train an acoustic model?

How much transcribed data do I need? How much training and testing data should I prepare?

--- (Edited on 3/1/2008 5:53 pm [GMT-0600] by nestea247) ---

Re: Training speaker dependent models from scratch
User: nsh
Date: 3/2/2008 5:22 am
Views: 266
Rating: 22

> How should I start (from scratch)?

Download sphinx3, try it. Then collect some recordings for adaptation (not more than half an hour).  Collect a lot of texts from your domain for language model (more than 50 Mb at least).

Train language model with cmuclmtk or use online tool:

http://www.speech.cs.cmu.edu/tools/lmtool.html

Adapt wsj to your speaker with MLLR:

http://www.speech.cs.cmu.edu/cmusphinx/moinmoin/AcousticModelAdaptation

Use both adapted model and  language model with sphinx3

 

--- (Edited on 3/2/2008 5:22 am [GMT-0600] by nsh) ---

Re: Training speaker dependent models from scratch
User: Visitor
Date: 3/4/2008 3:22 am
Views: 235
Rating: 19

Thanks for your pointers. They are very helpful. I'm still trying to go through the tutorial and compile sphinx3.

Just curious, why did you suggest sphinx3 and not julius? Is it because sphinx3 has an existing wsj model and julius has none yet? I don't know, but I find that julius is a cleaner and easier to understand system than sphinx3.

Btw, does anyone know when will voxforge's 1.0 model be released? 

--- (Edited on 3/4/2008 3:22 am [GMT-0600] by Visitor) ---

Re: Training speaker dependent models from scratch
User: kmaclean
Date: 3/4/2008 12:16 pm
Views: 1086
Rating: 21

Hi nestea247,

>does anyone know when will voxforge's 1.0 model be released?

when we get 140 hours of English Speech  :)

See the metrics page to see how much further we have to go.

Ken 

--- (Edited on 3/4/2008 1:16 pm [GMT-0500] by kmaclean) ---

PreviousNext