Mistral AI launches Voxtral TTS, a multilingual speech synthesis model

Mistral AI announced Voxtral TTS, its first text-to-speech model with 4B parameters, offering state-of-the-art multilingual voice generation. The model supports 9 languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic. It features low latency (70ms model latency), zero-shot cross-lingual voice adaptation, and contextual understanding for emotional expressiveness. Voice adaptation requires as little as 3 seconds of reference audio. Human evaluations showed superior naturalness compared to ElevenLabs Flash v2.5 while matching ElevenLabs v3 quality. The architecture includes a 3.4B transformer decoder backbone, 390M flow-matching acoustic transformer, and 300M neural audio codec. Available via API at $0.016 per 1k characters and in Mistral Studio. Part of the model weights are open on Hugging Face under CC BY NC 4.0.