how to use near field data trained acoustic model to fit far field audio?

[qin dong]

Hi everyone,

With 300 hours near field data trained, I've gotten a relative good WER (about 18%) with TDNN (nnet3) in my test set, which is also near field data. 

However, in real application, the data is mostly far field. 

I wanna know, in this case, what can I do to make the near field acoustic model fit the far field data better ?  

Is the beamforming in the front the best way to do this? Or the far field data to be trained is the only way? Or can I just simulate the near field data to far field?

Many thanks

 

[Daniel Povey]

You could look at what we did in the aspire recipe, where we added
reverberation and noise to simulate far-field data.
I think Vijay is working on making an easier-to-use version of that,
but for now, that's the only reververation recipe we have I think.

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.

[qin dong]

It is very helpful to me. Thank you very much for ur advice.

在 2016年4月29日星期五 UTC+8上午11:51:49，Dan Povey写道：

[Vijayaditya Peddinti]

You could also take a look at the reverb2014 task. It is the reverb2014 challenge system. A lot of teams participated in that challenge and the training data is simulated reverb data.

Vijay