Mistral's new Voxtral TTS clones voices in 3 seconds

Voxtral TTS is Mistral AI's new 4B-parameter text-to-speech model, built specifically to give real-time voice agents a fast, emotionally expressive voice.

It packs frontier voice quality into a compact footprint. The model runs with a 70ms latency on typical inputs and can clone an authentic-sounding voice using just 3 seconds of reference audio.

What it can do

Voxtral TTS natively covers 9 languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic, along with diverse regional accents.

Instead of robotic speech synthesis, the model picks up on contextual nuances like humor, happiness, or sarcasm. It preserves natural pauses, inflection, and subtle vocal quirks that make conversation feel genuine.

It also performs zero-shot cross-lingual voice cloning out of the box. Feed it a short French audio sample and an English sentence, and it will speak fluent English with a natural French accent—ideal for speech-to-speech translation pipelines.

The model generates up to two minutes of continuous audio natively, while the hosted API handles long-form generation seamlessly.

Under the hood and benchmarks

Under the hood, Voxtral TTS is an autoregressive flow-matching model built on top of Ministral 3B.

The architecture splits into three main components: a 3.4B-parameter transformer decoder backbone, a 390M flow-matching acoustic transformer, and a custom 300M neural audio codec operating at a 12.5Hz frame rate.

Speed is its biggest selling point for conversational agents. For a 10-second reference sample and 500 characters of text, Voxtral clocks a 70ms model latency with a 9.7x real-time factor.

In side-by-side human evaluations by native speakers, Voxtral TTS beat ElevenLabs Flash v2.5 on naturalness and acoustic similarity in zero-shot cloning, while matching the quality tier of ElevenLabs v3.

How to get hands-on

If you want to host it locally, Mistral released the model weights with reference voices on Hugging Face under a CC BY-NC 4.0 license.

For a zero-setup test, you can play with preset voices or record your own clip inside the Mistral Studio playground and Le Chat.

For production builders, the model is available right now via API, priced at $0.016 per 1,000 characters.

Why it matters

Anyone building conversational voice agents should give this a serious look. Developers have long been caught between pricey proprietary APIs with higher latency and clumsy open-source alternatives.

Voxtral TTS proves you do not need a massive model to achieve top-tier voice synthesis. Delivering expressive, multi-lingual audio in 70ms changes the cost-performance math for live customer support and interactive assistants.

The main caveat is the CC BY-NC 4.0 license on the open weights, meaning commercial deployments will either need Mistral's enterprise API or a custom licensing deal.