Voxtral TTS Models

Course updated November 3, 2026 for Voxtral TTS 3, based on the Migrate from Voxtral 2 guide and Mistral’s Text to Speech pages: voxtral-mini-tts-2603, deprecated October 31, 2026 and retired April 30, 2027, is replaced in examples by voxtral-mini-tts-3-0. Voxtral 2 predefined voices (gb_jane_neutral…) remain supported; Voxtral 3 adds [emotion=…] and [pace=…] tags in text, the expressiveness setting, custom pronunciation (phoneme_hints), and 35 languages. Mistral recommends re-measuring latency after migration. Sources: Migrate from Voxtral 2 · Speech generation.

Voxtral Mini TTS: The Reference Model

The synthesis model in the Voxtral family is Voxtral TTS 3, identified in the API as voxtral-mini-tts-3-0 (alias voxtral-mini-tts-3). This is the slug you will write in every API call: memorize it, as it is the only string in this course you need to know by heart. The model is designed to offer an optimal balance between voice quality, latency, and cost, which explains why it covers nearly all professional use cases on its own.

Voxtral Mini TTS is a voice foundation model. This distinction is not cosmetic: older TTS systems chopped text into phonemes and concatenated them from a sound bank, resulting in choppy voices where every syllable seemed independent of the previous one. Voxtral, by contrast, generates speech by understanding the semantic context of the text. On a sentence like “He didn’t say that,” the model knows where to place emphasis because it has interpreted the sentence structure, producing natural prosody with consistent intonation from start to finish.

Technical Specifications

The voxtral-mini-tts-3-0 model features the following specifications:

Those 90 milliseconds deserve a second thought: this is the delay before the first audio fragment becomes available, not the total generation duration. In a conversational assistant, this specific figure determines whether the exchange feels natural or laborious.

Pre-integrated Professional Voices

In addition to voice cloning, Mistral AI offers a library of professional voices ready for immediate use. These voices were recorded in a studio and optimized for various use cases. They provide invaluable service at the start of a project: if you want to verify that your pipeline works before even organizing a recording session with your chosen voice actor, a pre-integrated voice lets you test everything the same day.

In Mistral Studio, the web interface, you can explore these voices and filter them by three criteria: gender (male or female), the voice’s primary language, and emotion, meaning the emotional tone—neutral, enthusiastic, calm, and so on. Each is accessible via its voice_id in the API, exactly like a voice you cloned yourself: your code does not change a single line when you switch from the demo voice to the final voice.

Open Weights: Self-Hosting Options

The previous generation, Voxtral TTS (voxtral-mini-tts-2603), was released with its weights—a 4-billion-parameter model on Hugging Face—under the CC BY-NC 4.0 license. This is not an open source license in the usual sense: it prohibits commercial use. It allows experimentation and prototyping on your own infrastructure without any voice samples leaving your network, but a commercial product must use the API or reach an agreement with Mistral AI. As of today, the Voxtral TTS 3 documentation does not mention any open weights.

For organizations subject to confidentiality constraints—such as a healthcare provider or a government agency—the question therefore arises in two parts: is the use commercial, and does the open model cover the required languages and quality? If not, you must examine the API’s contractual commitments (hosting, data retention).

Usage via the Mistral API

To use Voxtral Mini TTS via the API, you must first initialize the client with your API key:

from mistralai.client import Mistral

client = Mistral(api_key="your-api-key")

The model is then referenced by its identifier voxtral-mini-tts-3-0 in every generation call. You will see in upcoming lessons how to create voices and generate speech; this client object will be the starting point for all following examples.

If you prefer a no-code approach, Mistral Studio offers a graphical interface accessible at studio.mistral.ai. There, you can listen to available professional voices, record your own voice directly from the browser, test voice generation with different texts, and adjust parameters—language, emotion, format. It is the ideal prototyping tool: validate in ten minutes that a voice suits the project, then move to code once the decision is made.

Roadmap Evolution

Mistral AI has announced that other models will join the Voxtral family: lighter models for embedded and mobile applications, more powerful models for expressive voice generation, and continuous improvements to multilingual quality. Therefore, design your integrations by isolating the model name in a configuration variable rather than scattering it throughout your code—the day a successor arrives, migration will take just one line. Until then, voxtral-mini-tts-3-0 remains the reference and covers the vast majority of professional use cases.

Key Takeaways

• The primary model is voxtral-mini-tts-3-0, optimized for quality and latency
• Pre-integrated professional voices are available to get started without an audio sample
• The previous model is open weights and can be self-hosted (non-commercial)
• Mistral Studio offers a graphical interface for no-code testing
• The Mistral API and Python SDK are the entry points for integration