support batching for kaldi compliant feature extraction functions

[saurabh-kataria]

🚀 Feature

batch dimension should be supported for kaldi complaint functions, for example, in torchaudio.compliance.kaldi.fbank

Motivation

Computation on GPU and use batches is essential

[mthrok]

Hi @saurabh-kataria

Thanks for submitting the feature request. This sounds reasonable request, but requires careful design.
As a starting point, can you describe how you would feed batched tensor?
like shape of the input tensor (what dimensions they represent) and how the function signature would change (if change is required).

[echocatzh]

I also encountered this problem with compliance.kaldi.fbank. I hope torchaudio can add batch processing operations, such as limiting the input dimension to 3 dimensions, [batch, channel, samples], or adding a batch_first option, because when working on asr or kws, the batchsize is usually very Large, if only one audio can be processed at a time, will the efficiency be reduced?