Multi Accent S2T Model

[bharat_patidar]

Hi everyone,

Can we use Deepspeech to build a multi accent S2T model?
example: I want to build an ASR model that can transcribe US and Indian accent English language. I have 3-4k hours of labelled data for each accent. Is it possible to build a single model for these accents which can give me good enough WER(<15)?

Any suggestion would be of great help!

Thanks

[othiele (Olaf Thiele)]

This is just my view, if any of you have another view. Please, go ahead

@bharat_patidar

Can we use Deepspeech to build a multi accent S2T model?

Yes, with a lot of data as you need to “store” more information.

@bharat_patidar

Is it possible to build a single model for these accents which can give me good enough WER(<15)?

The current model should be fine for US English, or isn’t it? You therefore want to finetune it for Indian English without loosing too much information. Use a low error rate and have a good test set that measures both which you can use to test how many epochs, what error rate and dropout you need to do that.