Fix docstring of sliding_window_cmn

[jackyguo624]

audio/torchaudio/functional/functional.py

Line 947 in ea85794

specgram (Tensor): Tensor of audio of dimension (..., freq, time)

the input specgram shoud be in shape ( ..., time, freq), instead of (..., freq, time) as doc says.

Right?

[mthrok]

Hi @jackyguo624

Thanks for the report. I think you are right about it and we should fix it.

[mthrok]

CC @jcaw Sorry we missed it in the review process.

[yoyololicon]

@jackyguo624
Hmm, I think the docstring is right, (..., freq, time) means you can drop the batch dimension, directly feed a matrix into it.

audio/torchaudio/functional/functional.py

Line 1022 in ea85794

if len(input_shape) == 2:

This line handle when input is a 2-dimensional matrix.

[mthrok]

@jackyguo624
Hmm, I think the docstring is right, (..., freq, time) means you can drop the batch dimension, directly feed a matrix into it.

audio/torchaudio/functional/functional.py

Line 1022 in ea85794

if len(input_shape) == 2:

This line handle when input is a 2-dimensional matrix.

@yoyololicon

Well, the batch/channel dimension is fine. I think the issue is about the order of the last two dimensions. The expected order is (..., time, freq) whereas the docstring is (..., freq, time)

[yoyololicon]

Oh, I see it, nevermind.😅

[jcaw]

Spectrograms returned by torchaudio do have the order (..., freq, time) though. Should sliding_window_cmn be operating on a different order? That seems like a bug.

[mthrok]

Spectrograms returned by torchaudio do have the order (..., freq, time) though. Should sliding_window_cmn be operating on a different order? That seems like a bug.

This is the inconsistent design principles between the libraries torchaudio followed. sliding_window_cmn is from Kaldi while spectrogram is from another different source. We could think of changing it but it will be a bc-breaking change, so it needs a separate discussion. For now, let's make the docstring of sliding_window_cmn accurately reflect the operation.

[mohamedirfansh]

Hello, can I work on this and fix this issue?

[mthrok]

Thanks. The docstring itself has changed. I'll think of what we can do about the inconsistent design. There are bunch of inconsistent interface design around Kaldi-originated functions.

[mthrok]

Thanks. The docstring itself has changed. I'll think of what we can do about the inconsistent design. There are bunch of inconsistent interface design around Kaldi-originated functions.

and if you have any suggestion, we are happy to hear.