French AI company Mistral has released Voxtral TTS, a lightweight open-source text-to-speech model with just 4 billion parameters that can run on consumer hardware from laptops to smartphones. Built on the Ministral 3B architecture, the model supports nine languages including English, French, German, Spanish, and Arabic, and can clone any voice from as little as three seconds of reference audio.
Human evaluations show Voxtral TTS matches ElevenLabs Flash v2.5 in naturalness and performs on par with the larger v3 model in lifelike interactions. The model delivers a time-to-first-audio of 90 milliseconds and renders a 10-second clip in roughly 1.6 seconds, making it fast enough for real-time voice agents, dubbing, and live translation.
Voxtral TTS is available under a Creative Commons license on Hugging Face, as well as through Mistral Studio and Le Chat. The release puts Mistral in direct competition with ElevenLabs, Deepgram, and OpenAI in the rapidly growing voice AI market.
