how to use Cuda for mfcc feat processing

[pawel-rozwoda]

Is any possibilty to generate feats like mfcc using Cuda in pyKaldi ?

Here is some snippet

logmel_opts = FbankOptions()
logmel_opts.frame_opts.samp_freq = SAMPLERATE
"""N_FILTERS = 40"""
logmel_opts.mel_opts.num_bins = LOG_MEL_FILTERS
fbank = Fbank(logmel_opts)

sf = logmel_opts.frame_opts.samp_freq
channel = Vector(signal)
cuda_vector = CuVector.from_vector(channel)
feats = fbank.compute_features(cuda_vector, sf, 1.0)

code above generate error like this:

Traceback (most recent call last):
  File "vox_generate.py", line 53, in <module>
    feats_0 = torch.Tensor(logmel_feats(signal=ch_0_filtered))
  File "/home/prozwoda/diarization/src/kaldi_feats.py", line 29, in logmel_feats
    feats = fbank.compute_features(cuda_vector, sf, 1.0)
TypeError: compute_features() argument wave is not valid for ::kaldi::VectorBase<float> (kaldi.cudamatrix._cu_vector.CuVector instance given): expecting kaldi.matrix._kaldi_vector.VectorBase instance, got kaldi.cudamatrix._cu_vector.CuVector instance

When I pass channel which is of type Vector into compute_features everything work fine, but on CPU.