Add support for 24 kHz sampling rate

[ChristianGeng]

Description

We should add support for 24 kHz (24000 Hz) sampling rate across the library. This sampling rate is commonly used in various audio applications and would be a valuable addition to our supported rates.

Current State

Currently supported sampling rates:

• 8000 Hz
• 16000 Hz
• 22500 Hz
• 44100 Hz
• 48000 Hz

Proposed Changes

Add 24000 Hz to:

• SAMPLING_RATES list in audb/core/define.py
• Test cases in tests/test_convert.py
• Documentation in docs/load.rst
• doctests in core

Motivation

24 kHz is a common sampling rate used in:

• Speech recognition systems
• Voice assistants
• Telephony applications
• Audio compression formats optimized for voice

Implementation Notes

The changes should be straightforward as we already have the infrastructure to support different sampling rates. We just need to add the new rate to the existing list and update relevant tests and documentation.

Testing

We should verify:

• Loading and converting audio files to 24 kHz
• Converting from 24 kHz to other rates
• All existing tests pass with the new rate

Documentation

Update the sampling rate list in the documentation to include 24 kHz while maintaining the ascending order.

/label enhancement

[hagenw]

I also wonder, why we support 22,500 Hz. This seems like a typo to me, as a common sampling rate is 22,050 Hz, e.g. as used in CSS10 and LJSpeech. See also https://github.com/audiojs/sample-rate.

As 24,000 Hz is close to 22,050 Hz, do you have examples for datasets stored with 24,000 Hz?

[hagenw]

I found a few examples myself. Seems to be that 24,000 Hz is used by TTS datasets, e.g. https://www.arxiv.org/abs/2408.06227.

[ChristianGeng]

44100 / 2 is indeed 22050. 22500 is indeed probably not so relevant :-)

…

On Tue, 19 Nov 2024, 08:34 Hagen Wierstorf, ***@***.***> wrote: I also wonder, why we support 22,500 Hz. This seems like a typo to me, as a common sampling rate is 22,050 Hz, e.g. as used in CSS10 and LJSpeech. See also https://github.com/audiojs/sample-rate. As 24,000 Hz is close to 22,050 Hz, do you have examples for datasets stored with 24,000 Hz? — Reply to this email directly, view it on GitHub <#466 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADMZTQ5LC6NIVHRTWKCWPO32BLSW3AVCNFSM6AAAAABSAXJMKSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOBUHA4TOMZRGU> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[ChristianGeng]

Yes they are used in TTS a lot but I think this is part of a more general move from 16k to 24k. There seems to be lots of use of neural codecs these days for other purposes as well. E.g. https://github.com/coqui-ai/TTS use it as well aot

…

On Tue, 19 Nov 2024, 08:41 Hagen Wierstorf, ***@***.***> wrote: I found a few examples myself. Seems to be that 24,000 Hz is used by TTS datasets, e.g. https://www.arxiv.org/abs/2408.06227. — Reply to this email directly, view it on GitHub <#466 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADMZTQ64W37CFOMU7FFBLS32BLTRHAVCNFSM6AAAAABSAXJMKSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOBUHEYTINJZGU> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[hagenw]

OK, then let's first add support for 24,000 Hz (as started in #467). And afterwards work on fixing 22,050 Hz as described in #468.

[ChristianGeng]

OK, then let's first add support for 24,000 Hz (as started in #467). And afterwards work on fixing 22,050 Hz as described in #468.

I assigned the review to you.