Voxtral TTS
Mistral AI
Voxtral TTS is Mistral's open-weights text-to-speech model for voice agents, released with BF16 weights and 20 reference voices. It covers nine languages (English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic and Hindi), does zero-shot voice cloning with no transcript required for the voice prompt, and streams with roughly 90 ms time-to-first-audio. Output is 24 kHz audio in WAV, PCM, FLAC, MP3, AAC or Opus. The weights inherit the CC BY-NC 4.0 licence of the reference voices, so self-hosting is non-commercial.
Strengths
- Zero-shot voice cloning with no transcript required for the voice prompt
- 20 preset reference voices shipped with the BF16 weights, plus adaptation to new voices
- Nine languages with dialect support: English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic and Hindi
- Streaming and batch inference with roughly 90 ms time-to-first-audio; the vendor measures 70 ms latency at concurrency 1 on a single H200
- 24 kHz audio output in WAV, PCM, FLAC, MP3, AAC and Opus formats
- Runs on a single GPU with at least 16 GB of memory in BF16, with production support in vLLM-Omni 0.18.0
- Exposed through Mistral's /v1/audio/speech endpoint for both text-to-speech and voice cloning
Caveats
- The weights are CC BY-NC 4.0, inherited from the reference voices, so self-hosted commercial use is not licensed; commercial deployments must use the paid API.
- Nine languages only: no Chinese, Japanese, Korean or Russian.
- The vendor publishes no speech-quality benchmarks such as intelligibility, word error rate or listener preference. The only table on the card reports latency, real-time factor and throughput, and notes that the reference script uses an inverted RTF formula that had to be converted back.
- Zero-shot voice cloning is an obvious impersonation risk, and the card's only mitigation is a 'Responsible Use' warning placing the compliance burden on the user.
- Neither a context window nor a maximum output length is published, and the API's output pricing of $16 per million characters is a meaningful cost for long-form narration.
- Vendor
- Mistral AI
- License
- CC-BY-NC-4.0
- Release Date
- 2026-03-23
- Modalities
- textaudio
Capabilities
Resources
Pricing
Mistral's model page prices the hosted model at $0 per million input characters and $16 per million output characters (EUR 10 in the EUR view). The open weights at mistralai/Voxtral-4B-TTS-2603 are CC BY-NC 4.0, so self-hosting is restricted to non-commercial use; commercial use runs through the hosted API.