OPUS files and ground truth for sampling rate

[hagenw]

The GigaSpeech dataset contains audio in opus files stored at 16,000 Hz.

I attached the file POD0000002525.opus to this issue.

>>> import audiofile
>>> file = "POD0000002525.opus"
>>> audiofile.sampling_rate(file)
16000
>>> audiofile.duration(file, sloppy=True)
536.144

But when reading the file in the usual way, the sampling rate no longer matches:

>>> signal, sampling_rate = audiofile.read(file)
>>> signal.shape
(25734790,)
>>> sampling_rate
48000

This also affects audiofile.samples() and audiofile.duration() as both uses audiofile.convert_to_wav(), which relies on audiofile.read() to get the ground truth for sampling rate:

>>> audiofile.samples(file)
25734790
>>> audiofile.duration(file)
1608.424375

This could be fixed by relying on audiofile.sampling_rate() inside audiofile.convert_to_wav(), but maybe we should also check, why the opus file behaves like this.

POD0000002525.zip

[hagenw]

ffmpeg always returns opus files wit a sampling rate of 48000 Hz, if not otherwise specified. This is known and marked as not to be fixed, see https://trac.ffmpeg.org/ticket/5240. Which means in order to fix it we need to provide ffmpeg with the desired sampling rate when converting an opus file to wav.