finetune aspire acoustic model

[ma...@nyu.edu]

I am getting some encouraging results on my data using the aspire acoustic model with my own LM  and  now wanted to fine tune the acoustic model (I had successfully done this in another context using a librispeech model I had built myself). In that model, it expected a GMM and also ran fmllr as a pre-processing step on the acoustic data. I notice that the pretrained aspire model does not come with a set of GMMs or fMLLR. Is this correct? And if so, can someone suggest the best way to fine-tune the aspire AM? 

Best

Michael 

[nshm...@gmail.com]

Btw, ASPIRE model has very old neural network architecture and not very accurate due to that.  You can probably try more modern Appen model from https://github.com/Appen/UHV-OTS-Speech instead. We adapted it for Vosk recently.

[Nathan Lindle]

Is the xconfig for that model available?  I'd be interested in checking out the differences.

Thanks

[nshm...@gmail.com]

You can run nnet3-info on final.mdl and check. It is more or less common tdnnf architecture like the one in librispeech recipe.