Voxtral TTS 4B

Tested as the voice. It sounds better than every other local voice measured here, but it generates slower than real time, so playback stops in the middle of sentences.

Voxtral TTS was released in March 2026. It speaks 9 languages including French, and clones a voice from 3 seconds of audio. It has no ONNX export, so it runs in Python through mlx-audio on Apple Silicon, not in Node. With mlx-community/Voxtral-4B-TTS-2603-mlx-4bit in streaming mode, it loads in 2.4 seconds and returns its first chunk of audio after 1.0 to 1.5 seconds, then produces about two thirds of a second of speech per second of computation, which is slower than real time.

That is too slow: the audio arrives slower than it plays, so playback stops in the middle of sentences. The same mlx-audio server now runs Qwen3-TTS, which also speaks French and generates faster than real time. Piper answers in 390 ms on the same machine and generates much faster than real time, so every sentence plays without a pause. Voxtral TTS sounds better than every other local voice, and on a machine two or three times faster it would be the best local option.

Voxtral TTS is also under a CC BY-NC license, so commercial use has to go through the Mistral API, not these weights.