Speaker adaptation of acoustic model for Indian accent kaldi ASR

[valluri....@impartus.com]

I am working on getting speech recognition for Indian accent speakers. Presently, I am using the online nnet2 decoding tool of Kaldi ASR. The tool is working well when the speaker has good english pronunciation. But, it is failing when the speaker speaks in a accent different from the US english accent.

So, can anyone please suggest any procedure for speaker adaptation of acoustic or neural network model using Kaldi ASR?

[Daniel Povey]

You need indian-accent training data if you want to do this. For
regular GMM models you might get something out of MAP adaptation (see
steps/train_map.sh), but for neural net models this does not work: you
would need to get a substantial amount of transcribed Indian-accented
data (preferably hundreds to thousands of hours) to train your system
on.

Dan

> --
> You received this message because you are subscribed to the Google Groups
> "kaldi-help" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to kaldi-help+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Valluri Saikiran]

Thank you for your reply.

I am presently using the open source neural network model in http://www.kaldi-asr.org/downloads/build/2/sandbox/online/egs/fisher_english/s5/. 

It worked well for good english pronunciation.

Can you please suggest a way by which I can adapt the model to an Indian speaker, for whom I have small amount of transcribed data , say 10 utterances each of average duration 10 sec.

Or else is it possible to do per speaker adatation of the model using any unsupervised learning from some hours of audio data of the speaker.

You received this message because you are subscribed to a topic in the Google Groups "kaldi-help" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/kaldi-help/7UsKhLtBXBM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-help+...@googlegroups.com.