Resample to 24 kHz
Drop an audio file and export it at 24 kHz as a WAV. The rate carries everything up to 12 kHz and sits at exactly half of 48 kHz.
An exact halving
48000 / 24000 is 2. Every output sample falls precisely on an input sample, so the resampler discards samples rather than interpolating between them. There is no fractional arithmetic and no interpolation error at all.
That makes it the cleanest downward step available from a 48 kHz source, which is not true of 44.1 kHz or 32 kHz, where the grids never quite line up.
Where the rate is used
Speech synthesis models commonly generate at 24 kHz. It gives a voice enough bandwidth to sound natural, including most of the sibilance that 16 kHz cuts short, while staying well below the data cost of a full music rate. Voice assistants and audiobook pipelines use it for the same reason.
The 12 kHz ceiling
Half the sample rate sets the limit. 12 kHz is generous for speech and audibly short for music: cymbals lose their top and a mix sounds closed in. Content above the limit is filtered out before the rate drops so it cannot alias back into the audible range.
Good for playback, not for recognition
Speech models are trained at 16 kHz and resample whatever they are given. A 24 kHz file is downsampled on the way in, so the extra bandwidth buys nothing for transcription. Use 16 kHz there and keep 24 kHz for audio people will actually listen to.
For music use 44.1 kHz or 48 kHz. For a mono downmix or another output format, use the full Sample Rate Converter.