Voxtral Mini 3B

Tested as the transcription, instead of Whisper. It transcribes French more accurately, but it takes 35 times longer per utterance, which is too slow for a call.

Voxtral Mini transcribes, translates and answers questions about audio. There is a community ONNX export, so it runs in Node through Transformers.js, like Whisper. examples/advanced/server/src/tests/voxtral-stt.ts runs it on French sentences spoken by Piper, with weights in q4 on CPU:

Model	Per utterance	Word errors
Voxtral Mini 3B	~16600 ms	17 %
Whisper base	~470 ms	34 %
Whisper french	~1200 ms	11 %

Voxtral is more accurate than Whisper base, but 35 times slower, because it transcribes by generating tokens with a 3 billion parameter model, where Whisper runs an encoder and a small decoder. Whisper french is more accurate than both and takes a fourteenth of Voxtral’s time. The weights take 3 GB in q4. Transformers.js fails when it downloads files that large, so download them with hf download first.