Speech Recognition Engines

Flat
how much training required for julius
User: sagar
Date: 9/21/2009 1:29 am
Views: 8461
Rating: 1
Hello,

I want to train julius for 20 words only.
1) How much training is sufficient, if I want to generate speaker independent model?
2) How can I train julius for noisy environment?
3) How to specify filler words in grammer file?
4) Is there any way to reject unknown words?
5) How can I generate N-gram file?
6) How can I use -mapunk option available in release 4.1.2 of julius?
7) Which julius options help me to improve accuracy?
8) How can I generate GMM definition file required for -gmm option?

--- (Edited on 9/21/2009 1:29 am [GMT-0500] by Visitor) ---

Re: how much training required for julius
User: kmaclean
Date: 9/21/2009 2:35 pm
Views: 243
Rating: 2

Hi sagar,

>I want to train julius for 20 words only.

>1) How much training is sufficient, if I want to generate speaker

>independent model?

Depends on many things... for example: the type of application you are planning to create, user accents/dialects, how different are words to be recognized (if they are similar, then more speech required)...

The current Sphinx3 acoustic model uses 140 hours of speech.

You should be able to get away with much less if you are only looking at 20 words - I would *guess* 5-10 hours from at least 100-200 different speakers. 

>2) How can I train julius for noisy environment?

Train acoustic models with speech from the same noisy environment that your target application will operate in:

see these posts for more info:

>3) How to specify filler words in grammer file?

Train your acoustic model with non-speech sounds, and include such non-speech sounds in your pronunciation dictionary and grammar.

>4) Is there any way to reject unknown words?

see my second post in this thread: Re: One word grammar, always recognized?

>5) How can I generate N-gram file?

6) How can I use -mapunk option available in release 4.1.2 of julius?

don't know... best to ask on the Julius forum...

7) Which julius options help me to improve accuracy?

Read through the Julius config file - lots of hints there...

Then trial and error.  Very important to have test files so you can automate your tests, and know when performance improves.

I have used these parameters (from Julian jconf - should be similar in Julius 4.*):

######################################################################
#### Language Model
######################################################################
##
## word insertion penalty
##
-penalty1 5.0        # first pass
-penalty2 20.0        # second pass

##
## For insertion of context-free short-term inter-word pauses between words
##  (multi-path version only)
##
-iwsp            # append a skippable sp model at all word ends
-iwsppenalty -70.0    # transition penalty for the appenede sp models

8) How can I generate GMM definition file required for -gmm option?

best to ask on the Julius forum... likely an HTK tool that creates this.

Ken

--- (Edited on 9/21/2009 3:35 pm [GMT-0400] by kmaclean) ---

Re: how much training required for julius
User: Alex
Date: 10/12/2010 6:42 am
Views: 69
Rating: 2

Hi!

I am planing to perform specific project, a model that will recognize just one speaker.


Basically the software will be trained by me and I will be the only person using it.

 

In such case what kind of training is needed?

 

Thanks in advance, Alex

--- (Edited on 10/12/2010 6:42 am [GMT-0500] by Visitor) ---

Re: how much training required for julius
User: Visitor
Date: 10/16/2010 10:57 am
Views: 2180
Rating: 1

>In such case what kind of training is needed?

THe CMU Sphinx acoustic model training page provides some rules of thumb that should be applicable to Julius too:

  • 1 hour of recording for command and control for single speaker
  • 5 hour of recordings of 200 speakers for command and control for many speakers
  • 10 hours of recordings for single speaker dictation
  • 50 hours of recordings of 200 speakers for many speakers dictation

--- (Edited on 10/16/2010 10:57 am [GMT-0500] by Visitor) ---

PreviousNext