Mistral AI released Voxtral TTS on Thursday, a text-to-speech model that the Paris-based company claims outperforms ElevenLabs' offerings while providing enterprises full control over their voice AI infrastructure. Unlike competitors who operate API-based services, Mistral is releasing the complete model weights for free download.

The 3.4-billion-parameter model supports nine languages including English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic. It can run on consumer-grade hardware including laptops and smartphones, requiring approximately 3 gigabytes of RAM when optimized. The model achieves 90-millisecond time-to-first-audio response and generates speech at six times real-time speed.

In internal evaluations, Mistral reported that human listeners preferred Voxtral TTS over ElevenLabs Flash v2.5 62.8 percent of the time for standard voices and 69.9 percent for voice customization tasks. The model can adapt to custom voices using as little as five seconds of reference audio and demonstrates cross-lingual voice adaptation capabilities.

The release represents Mistral's broader strategy of providing enterprises with AI infrastructure they can own rather than rent through cloud APIs. Pierre Stock, Mistral's vice president of science, emphasized that voice recordings contain sensitive data that many compliance-focused industries prefer not to send to third-party services. The company has positioned itself as a European alternative to American AI providers, targeting the estimated $22 billion global voice AI market.

Voxtral TTS complements Mistral's existing audio offerings, including the recently released Voxtral Transcribe speech-to-text model. Together with the company's language models and enterprise platform services, these tools form what Mistral describes as a complete AI stack for voice applications. The model is available for testing through Mistral's API and can be downloaded for local deployment.