The Voxtral Story

Voxtral TTS is Mistral's first text-to-speech model, aiming for top-tier performance in multilingual voice generation. Designed at a 4B parameter scale, it emphasizes a cost structure suited for natural, reliable, and large-scale operation.

The core is not simple reading aloud but context understanding and speaker modeling. It reflects emotion, intonation, pauses, rhythm, and speech habits to create more human-like voices, and is designed so that companies can directly control their own voice AI stack.

Key features are as follows.

• Support for 9 languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic
• Support for various dialects and emotional expression
• Low-latency streaming with very short time-to-first-audio
• Custom voice adaptation possible with as little as 3 seconds of reference audio
• Can be tested directly in Mistral Studio

On the performance side, Mistral believes that human evaluation based on native speakers matters more than automated metrics. In comparative evaluations, Voxtral TTS showed superior naturalness compared to ElevenLabs Flash v2.5, maintained a similar level of TTFA, and was described as being on par in quality with ElevenLabs v3. In particular, in a zero-shot custom voice setting, when comparing naturalness, prosody appropriateness, and acoustic similarity to the original across representative voices in 9 languages, it showed an even larger quality gap.

The architecture is a transformer-based autoregressive flow-matching model, built on top of Ministral 3B. It consists of a 3.4B parameter transformer decoder backbone, a 390M flow-matching acoustic transformer, and a 300M neural audio codec. Voice prompts of 5 to 25 seconds in length are used, and after predicting a semantic token for each audio frame, the flow-matching transformer generates acoustic latents with 16 NFEs.

Low-latency streaming is also a strength. Based on a typical input of a 10-second voice sample and 500-character text, it presents 70ms model latency and roughly 9.7x RTF, with the model directly generating up to 2 minutes, while longer segments are handled via the API's smart interleaving.

A representative enterprise use case is customer service centers. Voxtral TTS can be paired with Voxtral Transcribe to complete a speech-to-speech pipeline, or combined with existing speech-to-text and LLM stacks to implement multilingual voice agents. The API price is $0.016 per 1,000 characters, and an open weights version is also provided on Hugging Face.