Mistral AI released Voxtral TTS on March 26, 2026, a text-to-speech model that the French company says can reproduce a speaker’s voice from only a few seconds of reference audio. The model is published with open weights, meaning developers can download and run it themselves rather than relying solely on Mistral’s hosted service.
Voxtral TTS is built on Ministral 3B, one of Mistral’s small language models, and has about 4 billion parameters in total including its audio components. Mistral says the model supports nine languages and is small enough to run on consumer hardware, including on devices rather than only in data centers.
Voice cloning in seconds
The feature Mistral emphasizes most is zero-shot voice cloning. According to the company, a sample of roughly five seconds is enough for the model to generate new speech in the same voice, without any fine-tuning. The company also highlights low latency, which matters for conversational applications where the system must reply quickly.
Mistral says that in its own human preference tests, listeners favored Voxtral TTS over ElevenLabs’ output in a majority of comparisons. That is a company-run evaluation; independent benchmarks were not available at launch.
Completing an audio stack
The release rounds out the Voxtral family, which began in 2025 with speech-to-text models for transcription and audio understanding. With both recognition and generation now covered, Mistral can offer a complete pipeline for voice agents on a single platform: listen to a user, reason with a language model, and reply in natural speech.
By releasing the weights openly, Mistral is taking the same approach it has used for many of its language models. The company argues that open weights let businesses adapt the model to their own needs, run it on their own infrastructure for privacy reasons and avoid per-request pricing. The model is also available through Mistral’s API for customers who prefer a hosted option.
Competitive pressure on speech providers
Voxtral TTS enters a market dominated by proprietary services from ElevenLabs, OpenAI, Google and Deepgram, among others. A capable open-weight alternative that can run locally puts pricing pressure on those providers, particularly for high-volume uses such as audiobooks, customer service and accessibility tools.
The ease of cloning voices from short samples also raises familiar concerns about impersonation and fraud. Mistral’s release materials describe the capability as a feature for personalization and localization, and the company’s terms of use prohibit cloning a voice without the speaker’s consent. How effectively such rules can be enforced for a model that anyone can download remains an open question for the industry as a whole.
