Voxtral TTS

Learning Objectives

• Understand what the Voxtral speech models do and how the March and October releases differ
• Know what the non-commercial license on the downloadable weights does and does not allow
• Evaluate when Voxtral fits a voice agent or narration project against commercial alternatives

What Is Voxtral TTS?

Voxtral TTS is Mistral AI's text-to-speech line. The first model, released on March 26, 2026, has 4 billion parameters, small enough to run on a single consumer GPU, with 20 preset voices across 9 languages and the ability to adapt to a new voice from a short audio reference.

On October 20, 2026 Mistral launched Voxtral 3, a speech family built for voice agents rather than a single model. It pairs a new text-to-speech model, Voxtral Mini TTS 3, with new transcription models and a developer kit that joins listening, reasoning, tool calls and speech into one live stream.

Voxtral 3: Built for Voice Agents

Mistral's pitch for Voxtral 3 is that a working voice agent has to listen, reason, act and speak at the same time, and that stitching separate products together for each step is where most projects break down. The family has three parts.

• Transcription. Voxtral Small Transcribe 3 and the smaller 4 billion parameter Voxtral Mini Transcribe 3 cover more than 50 languages, identify who said what, accept a list of key terms and names to improve accuracy, and detect when a speaker has finished. A realtime version targets under half a second of latency in more than 35 languages. Mini costs about a third of a cent per minute.
• Speech. Voxtral Mini TTS 3 adjusts emotion, pace and emphasis to fit the conversation, in more than 35 languages and 50 dialects, at 13 dollars per million characters. A Pro Voice option reproduces a specific speaker's character and style across languages from about 30 minutes of audio.
• The agent kit. A voice agent SDK runs listening, reasoning, tool calls and speaking as one real-time pipeline for phone and web, so an agent can look something up or book an appointment mid-conversation.

Everything runs on Mistral's EU-hosted API or on a customer's own infrastructure, with control over whether any audio is retained. That is the sovereignty argument Mistral makes across its whole product line, applied to voice.

Key Capabilities of the Original Model

Multilingual Support

The March model supports 9 languages:

• English, French, Spanish, German, Italian, Portuguese, Dutch, Russian, Chinese

Natural Prosody and Voice Adaptation

The model generates speech with natural rhythm, intonation and emphasis, and ships with 20 preset voices. Given a 10-second reference recording, it can adapt to a new voice.

• Contextual emphasis — stresses important words based on meaning
• Natural pauses — appropriate breathing and sentence breaks
• Expressive variation — adjusts tone for questions, statements, and exclamations

Runs on Consumer Hardware

At 4 billion parameters, the March model is designed to run locally:

• Runs on a single consumer GPU (NVIDIA RTX 3090 or equivalent)
• No cloud dependency required for inference
• Suitable for edge deployment and privacy-sensitive applications, within the non-commercial license

Pricing

Voxtral vs. Other TTS Solutions

Strengths

• Downloadable model for research and personal projects — the March weights run locally without API costs
• Consumer hardware — 4 billion parameters runs on a single GPU; no data center required
• One vendor for the whole voice agent — Voxtral 3 covers transcription, speech and the pipeline between them
• Low API cost — 13 dollars per million characters for Voxtral Mini TTS 3 undercuts several commercial alternatives, by Mistral's comparison
• European AI — built by Mistral AI in Paris, with an EU-hosted API and self-deployment for data sovereignty

Limitations and Considerations

• Non-commercial weights — the March model's CC BY-NC 4.0 license rules out using the free download in a commercial product; self-hosting for business means the paid route
• No open Voxtral 3 weights yet — the October models are API and self-deployment offerings, not a public download
• Vendor benchmarks — the Voxtral 3 comparisons come from Mistral, and independent rankings put some rivals ahead on transcription accuracy
• 9 languages on the March model — far fewer than Voxtral 3 or the large commercial platforms
• Young ecosystem — community tooling and fine-tuning resources are still thin compared with older speech platforms

Company Details

Related Tools

• ElevenLabs — Premium voice AI with cloning and real-time streaming
• Mistral Large 4 — Mistral's flagship language model
• Devstral — Mistral's coding-focused model

Key Takeaways

• Voxtral TTS was Mistral AI's first voice product, a 4 billion parameter model that runs on consumer hardware with 20 preset voices in 9 languages
• Its downloadable weights are licensed for non-commercial use only, so a business product needs Mistral's paid API or an agreement
• Voxtral 3, launched October 20, 2026, adds conversation-aware speech in 35-plus languages, transcription in 50-plus languages and a kit for building voice agents
• Voxtral Mini TTS 3 costs 13 dollars per million characters through Mistral's EU-hosted API, and the family can also be deployed on a customer's own servers
• Mistral's benchmarks favor its own models, and independent transcription rankings place Microsoft and Alibaba models ahead