Added GSM format support to save function

[imaginary-person]

SUMMARY

Added the changes listed in #1264 towards supporting GSM format in the save function.

CAVEATS

1. http://sox.sourceforge.net/soxformat.html doesn't list various parameters for gsm, so only a generic statement regarding GSM not supporting custom encodings/bit depth/compression has been added in the documentation, but I can revise it further, if required.

2. I also made some changes in the code-structure for readability, with most of them aimed at conformance with the Google style guide for C++.
   I even added curly braces to case blocks within switch statements, but I can roll back such changes, if required, as the Google style guide doesn't mandate one to do so.

3. I'm submitted this PR since the issue has a good first issue label, but if it's still being worked on by someone, please disregard this PR.

Thanks!

[mthrok]

Hi @imaginary-person

Thanks for the contribution. Let me talk with the original contributor first. If you would like, you can try HTK format too. It should be practically the same process.

[imaginary-person]

Hello @mthrok, SOX_ENCODING_HTK doesn't exist in the sox source-code. HTK encoding in sox actually uses SOX_ENCODING_SIGN2 under the hood, but since SOX_ENCODING_HTK can't simply be substituted with SOX_ENCODING_SIGN2, please advise if some sort of a workaround is possible. Basically, HTK is a format in sox but not an encoding. If I make Format::HTK correspond to SOX_ENCODING_SIGN2 in torchaudio/csrc/sox/utils.cpp, then the save tests for HTK fail:

save_test.py:250: in test_save_htk
    self.assert_save_consistency("htk", test_mode=test_mode)
save_test.py:140: in assert_save_consistency
    self.assertEqual(found, expected)
/home/pytorch/torch/testing/_internal/common_utils.py:1183: in assertEqual
    super().assertTrue(result, msg=self._get_assert_msg(msg, debug_msg=debug_msg))
E   AssertionError: False is not true : Tensors failed to compare as equal!Attempted to compare equality of tensors with different sizes. Got sizes torch.Size([1, 48000]) and torch.Size([1, 24000]).

Thank you!

[mthrok]

SOX_ENCODING_HTK doesn't exist in the sox source-code. HTK encoding in sox actually uses SOX_ENCODING_SIGN2 under the hood

HTK format uses signed-integer encoding. [code]

but since SOX_ENCODING_HTK can't simply be substituted with SOX_ENCODING_SIGN2

Why not? There are other formats that saves samples in raw format like wav, such as amb, sph. I expect that using SOX_ENCODING_SIGN2 should allow us to replicate the result sox produces.

If I make Format::HTK correspond to SOX_ENCODING_SIGN2 in torchaudio/csrc/sox/utils.cpp, then the save tests for HTK fail:

save_test.py:250: in test_save_htk
    self.assert_save_consistency("htk", test_mode=test_mode)
save_test.py:140: in assert_save_consistency
    self.assertEqual(found, expected)
/home/pytorch/torch/testing/_internal/common_utils.py:1183: in assertEqual
    super().assertTrue(result, msg=self._get_assert_msg(msg, debug_msg=debug_msg))
E   AssertionError: False is not true : Tensors failed to compare as equal!Attempted to compare equality of tensors with different sizes. Got sizes torch.Size([1, 48000]) and torch.Size([1, 24000]).

This is most likely the underlying encoder discards the channel information. http://sox.sourceforge.net/soxformat.html says,

.htk Single channel 16-bit PCM format used by HTK, a toolkit for building Hidden Markov Model speech processing tools.

Looking at the code, this format seems to have a specific header but the sample storage is same as WAVE format. Also the source code has the link for the references. (if you are interested in)

https://labrosa.ee.columbia.edu/doc/HTKBook21/node57.html
https://labrosa.ee.columbia.edu/doc/HTKBook21/node61.html
https://labrosa.ee.columbia.edu/doc/HTKBook21/node58.html

Anyways, you can provide num_channels parameter to change the number of channels used in test. Can you try that?

[imaginary-person]

Thanks for your help, @mthrok. I hadn't noticed that the number of channels used in that test were 2 by default.
By providing a num_channels parameter, tests for htk also passed, as you had expected.
I'll submit a separate PR for htk, so as to retain the existing code-style (I changed style in some files of this PR).

Why not? There are other formats that saves samples in raw format like wav, such as amb, sph. I expect that using SOX_ENCODING_SIGN2 should allow us to replicate the result sox produces.

Yes, I just meant that I'd have to make some other change as well, such as the num_channels one you suggested.

[imaginary-person]

Closed this PR as the original contributor submitted another one.