Resample to 16 kHz
Drop an audio file and export it at 16 kHz as a WAV. This is the rate speech recognition systems and voice models are built around.
Why speech pipelines standardised here
The acoustic cues that separate one phoneme from another live below 8 kHz. A 16 kHz sample rate carries everything up to that limit and nothing above it, which covers the useful information in speech while storing a third as much data as 48 kHz.
Models are trained on audio at this rate. Feed one a 44.1 or 48 kHz file and it resamples down before processing, so the extra samples are discarded on the way in. Converting first makes the file smaller and takes the resampling out of a step you cannot inspect.
What 8 kHz of bandwidth costs
Sibilance loses its top edge, cymbals and air go, and music sounds noticeably dull. Speech remains clearly intelligible, and that is the whole basis of the tradeoff.
The content above 8 kHz is filtered out before the rate changes. Without that filter those frequencies would fold back down into the speech band as aliased tones, which is far more damaging to recognition accuracy than losing the treble.
Smaller files, better encoding
For WAV and FLAC the saving is direct: a third of the sample rate means a third of the data. For MP3, AAC, and Opus the bitrate sets the size instead, so the rate change does not shrink the file. It gives the encoder less spectrum to spend those bits on, which is why 16 kHz mono holds up far better than 44.1 kHz stereo at a low bitrate.
Other delivery rates
Telephone-grade audio is 8 kHz. Music wants 44.1 kHz or 48 kHz. To downmix to mono at the same time, use the full Sample Rate Converter.