Hi chandu,

>but there is no difference in the final acoustic model.i mean, it was same

>as before.

Maybe your training speech is not that difficult for the training process to parse out the phones.  Timestamps can help with continuous speech where words in a sentence get mashed together to sound like one big continuous word, but if your training speech is relatively clear, then timestamps might not help.

>so can some one guide me how to use the time stamped label files

>generated by HSlab,

I am not sure if there is a switch in HTK to tell the training process to use timestamps, I think it picks it up automatically.   It has been a long time since I tried this approach (I was trying to training long speech samples - 20-30 minutes... I could not get it to work).  I had much better results segmenting speech into 15-25 word sentences and letting the HTK training process figure out where the phones begin and end.

Have you searched the HTK manual, or reviewed the HTK email archives?

Ken

Hello Mr.clean,

thanks for ur response and sorry for a delayed reply.

I searched a lot, but could not find any solution regarding the time stamped label files.then i gave it up and started slicing the actual speech into 5 to 10 words each.and recognitions got to be better.However for few of  the speakers,it is not working out.

today im happy to see ur post ,to know that u also followed the same procedure and achieved better results.Now im confident that my aproach is correct.

Can u clarify this doubt if possible.

I have created an acoustic model with say for example 50,000 sentences. and i have 1000 more sentences which i want to add to the existing acoustic model . is there any possibility for this, or do we have to train the 60,000 sentences together again?

Now im trying to create a Speaker Independent Acoustic model, by taking small sentences of continuous speech related to different speakers. if u have also worked on the same kind of stuff, can u plz share ur experience with me?

Thanks in advance,

Chandu ([email protected]).

>i have 1000 more sentences which i want to add to the existing acoustic

>model . is there any possibility for this, or do we have to train the 60,000

>sentences together again?

See the VoxForge adaptation tutorial

>Now im trying to create a Speaker Independent Acoustic model, [...]

>if u have also worked on the same kind of stuff, can u plz share ur

>experience with me?

the VoxForge acoustic model releases and nightly builds are speaker independent.  The script that is used to create these is similar to the approach taken in the VoxForge tutorial.

Hi Clean,
Thanks for your reply.

      I have tried these approaches.I have derived an Acoustic Model(as per adaptation tutorial) using latest VoxForge

Acoustic Model release and few samples of the speaker that i want.But the recognition results with julius are not that

satisfactory.

      Should there be any difference if we adapt countinuous speech(dictated at a speed)to the VoxForge Acoustic Models.I

hope that VoxForge Acoustic Model is built on a slowly dictated text like "dial One two three Steve Young"etc.

And does the language model have any impact on the quality of the recognition.

Waiting for your guidance,
Chandu

>Should there be any difference if we adapt countinuous speech(dictated at

>a speed)to the VoxForge Acoustic Models.

don't know, you'll have to experiment and see what happens...

Just make sure you are adapting with audio that is the same sample rate and bits per sample as the VoxForge acoustic model.

>And does the language model have any impact on the quality of the recognition.

Yes.  But I have very little experience with statistical language models, so I cannot give you much guidance in that regard...

>hope that VoxForge Acoustic Model is built on a slowly dictated text like

>"dial One two three Steve Young"etc.

depends on who is reading... you can listen to a sampling of the audio submissions to get an idea.

Just make sure you are adapting with audio that is the same sample rate and bits per sample as the VoxForge acoustic model.

iam sure about this. am using 16000 hz and 16 bits per sample in my audio samples. but still, normal AM with around 10 minutes of speech is performing atleast better then that adapted from voxforge acoustic models.

Yes.  But I have very little experience with statistical language models, so I cannot give you much guidance in that regard...

regarding this, iam using ngram-count tool to create an acoustic model. could you observe any better performance with the LM created with HTK tools?

and i somewhere read, that, herest is also used for adaptation.

since headapt is eliminated from the latest version of htk, herest must have got that feature with it. but i could not find any documentation on that. can you suggest something on this.

Thanks in advance,

Chandu.

@Mr Clean and Mr Chandu;

I'm student at Hanoi University of technologies. I have the same probleme like Mr. Chandou. I have to extract the time stamp of each phone for training. If you have the solution, contact with me by this email : [email protected]. Thanks so much for your help