Voxtral TTS

2026-03-23 · Publisher : Mistral AI · Domain : Artificial Intelligence

Mistral AI launches Voxtral TTS, its first text-to-speech model: 4 billion parameters, 9 languages, voice adaptation in a few seconds and open weights under the CC BY-NC 4.0 license.

Mistral AI launches Voxtral TTS, its first text-to-speech model: 4 billion parameters, 9 languages, voice adaptation in a few seconds and open weights under the CC BY-NC 4.0 license. ### What changes - Mistral's first text-to-speech model, compact and low-latency. - Expressive voices in 9 languages with dialects, adaptation of new voices (sample of 5 to 25 seconds). - Cross-lingual voice adaptation without prior training. - Up to 2 minutes of audio generated natively. ### Specifications - Size: about 4B (3.4B decoder, 390M acoustic transformer, 300M codec). - Languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic. - License: CC BY-NC 4.0 (weights on Hugging Face, non-commercial use; the API covers commercial use). - Endpoint: /v1/audio/speech. Availability: API, Mistral Studio, Le Chat, Hugging Face. - Announced latency: 70 ms for a typical 10-second sample. ### Pricing | Item | Price | |---|---| | API | 0.016 USD per 1,000 characters | ### Benchmarks | Measure | Result | |---|---| | Human preference on voice customization against ElevenLabs Flash v2.5 | 68.4% | Scores published by the vendor (human evaluation by native speakers). ### Key takeaways An open, low-cost voice building block complements Voxtral transcription to build voice agents that are fully controlled end to end.

Tags : LLM, Mistral, Voxtral, Audio, Synthèse vocale, Open Source