Fix return dtype in MVDR module

[nateanl]

Address #2375
The MVDR module internally transforms the dtype of complex tensors to torch.complex128 for computation and transforms it back to the original dtype before returning the Tensor. However, it didn't convert back successfully due to specgram_enhanced.to(dtype), which should be specgram_enhanced = specgram_enhanced.to(dtype). Fix it to make the output dtype consistent with original input.

[facebook-github-bot]

@nateanl has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[mthrok]

Can we have corresponding tests?

[facebook-github-bot]

@nateanl has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[mthrok]

Looks good. BTW can we move the implementation of MVDR into dedicated file, such as torchaudio/transforms/_mvdr.py?

[nateanl]

Looks good. BTW can we move the implementation of MVDR into dedicated file, such as torchaudio/transforms/_mvdr.py?

Sure, I prefer using _beamforming.py or _multi_channel.py as there are other beamforming methods liks GEV or WPD, and other multi-channel methods like WPE, IPD, etc.

Or you mean for this specific MVDR module for deprecation purpose?

[mthrok]

Looks good. BTW can we move the implementation of MVDR into dedicated file, such as torchaudio/transforms/_mvdr.py?

Sure, I prefer using _beamforming.py or _multi_channel.py as there are other beamforming methods liks GEV or WPD, and other multi-channel methods like WPE, IPD, etc.

Or you mean for this specific MVDR module for deprecation purpose?

What I meant was to put all the multi channel transforms to dedicated script.
_transforms.py is getting huge, but IIUC, multi-channel transforms are not logically connected to the rest of transforms.
Putting them in dedicated file should yield better decoupling.
You can pick the name of the file, as long as it represents the module well.