Mistral AI Releases Voxtral TTS: A 4B Open-Weight Streaming Speech Model for Low-Latency Multilingual Voice Generation

Mistral AI released Voxtral TTS, an open-weight text-to-speech model offered as part of its audio generation offerings.

Positioned as a direct competitor to proprietary voice APIs, Voxtral TTS is designed as a modular component for integration into real-time voice workflows.

The model is released under a CC BY-NC license.

Hybrid 4B-parameter architecture: 3.4B Transformer decoder (handles text understanding and predicts semantic speech representations), 390M flow-matching acoustic transformer (converts semantic representations into acoustic features), and 300M neural audio codec (maps acoustic features to a high-fidelity audio waveform).

Optimized for low-latency streaming inference; example model latency is 70ms for a typical 10-second voice sample from a 500-character input, suitable for conversational agents and real-time translation.