Loading sph audio files

[jiwidi]

Can audio.load load NIST sphere format (SPH) files?

Hi!

So I have tried to load sph files from tedlium dataset with simply:

torchaudio.load("pathtofile")

The load was successful but when saving it to a .wav file to check if the torchaudio read correctly the audio then I realized it was not working.

My question then is if this type of audio files is currently supported by torchaudio and if so, could you provide some link/code that refers how to? I haven't been able to find anything about torchaudio and sph files (but I know sox supports sph and allows to conversion into wav)

In case someone needs sph files here there are some sample ones https://filesamples.com/formats/sph. I got mines from the TEDLIUM dataset.

Thanks

[mthrok]

Hi @jiwidi

Thanks for reporting this.

Could you share the following things?

• Your environment: the output of python -m torch.utils.collect_env
• code snippet you used to save audio

The correctness of the save function in "sox" backend is never tested with "sph" format. We are moving away from "sox" backend but we do not support "sph" format in "sox_io" backend yet. Maybe we can add "sph" support to "sox_io" backend. -> #871

[jiwidi]

Yes @mthrok!

Environment on my macbook:

Collecting environment information...
PyTorch version: 1.7.0.dev20200807
Is debug build: No
CUDA used to build PyTorch: None

OS: Mac OSX 10.15.5
GCC version: (Homebrew GCC 4.9.4_2) 4.9.4
CMake version: version 3.14.0

Python version: 3.7
Is CUDA available: No
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA

Versions of relevant libraries:
[pip3] numpy==1.19.1
[pip3] torch==1.7.0.dev20200807
[conda] blas                      1.0                         mkl
[conda] mkl                       2019.4                      233
[conda] mkl-include               2020.1                      216
[conda] mkl-service               2.3.0            py37hfbe908c_0
[conda] mkl_fft                   1.1.0            py37hc64f4ea_0
[conda] mkl_random                1.1.1            py37h959d312_0
[conda] numpy                     1.19.1           py37h3b9f5b6_0
[conda] numpy-base                1.19.1           py37hcfb5961_0
[conda] pytorch                   1.7.0.dev20200807         py3.7_0    pytorch-nightly
[conda] torch                     1.6.0                    pypi_0    pypi

The code (with the first sample from the link in first message):

import torchaudio
wave,sample = torchaudio.load("sample1.sph")
torchaudio.save("sample1.wav",wave,sample)

I checked if the sample returned from the loading was the same specified as in the sph (44100) and they match.
I try open the wav file and its now a 2hour snoring version

[mthrok]

@jiwidi Thanks, and you are working on torchaudio master branch, right?

I added support for sph format in "sox_io" backend.
Can you try my branch mthrok/sph or cherry pick this commit mthrok@f78ebf9
and try "sox_io" backend?

import torchaudio
torchaudio.set_audio_backend("sox_io")
wave, sample = torchaudio.load("sample1.sph")
torchaudio.save("sample1.wav", wave, sample)

I believe this works, and if you can confirm I will merge it to master branch.

[jiwidi]

@mthrok
I'm working on master branch yes

I can confirm it works 👍 Thank you very much for the quick PR!

[mthrok]

@jiwidi Thanks for confirming. I merged the commit to master.