Introduction to Voxtral TTS

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.

Text-to-speech enters a new era

In March 2026, Mistral AI launched Voxtral TTS, a text-to-speech engine that pushes the boundaries of what was previously possible. To appreciate the leap, recall what the previous generation required: to get a convincing voice, you had to bring a voice actor into a studio, record hours of reading, and then train a model on that corpus. Voxtral takes a different approach with zero-shot voice cloning: just a few seconds of audio are enough to capture a voice’s timbre, rhythm, and nuances.

This breakthrough relies on a foundation model specifically trained for speech generation. Voxtral doesn’t just reproduce phonemes—it understands prosody, emotions, and linguistic subtleties across 35 languages and more than 50 dialects with Voxtral TTS 3, including French (the previous generation covered 9). You can hear the difference in an interrogative sentence or a parenthetical remark: where an older system would recite flatly, Voxtral places the accent correctly because it reads the entire sentence before speaking it.

What is zero-shot cloning?

The term zero-shot means the model doesn’t need retraining to reproduce a new voice. You provide a recording lasting a few seconds to a few minutes, in MP3 or WAV format, and Voxtral automatically captures vocal characteristics: timbre, intonation, rhythm, and accent. There’s no training phase, no extra compute cost, and no waiting period—the sample is analyzed at request time.

Consider a training manager who needs to voice an e-learning module. She records twenty seconds with a headset, creates her custom voice, and immediately has a narrator who sounds like her. She can also clone this voice on the fly directly in an API request without even saving it if the need is one-off. And when the module needs to be produced in English and Spanish, the same voice retains its identity across languages: the listener recognizes the person, not three different narrators.

Why Voxtral changes the game

Traditional TTS solutions—Google Cloud TTS, Amazon Polly, Azure Speech—offer pre-recorded voices of varying quality that you take as-is. Newer solutions like ElevenLabs offer voice cloning but remain proprietary and expensive: you rent access without ever controlling the engine.

Voxtral stands out on three mutually reinforcing fronts. The previous generation, Voxtral TTS, was released with its weights under the CC BY-NC 4.0 license: hostable on your own infrastructure for non-commercial use, which matters for a public service or healthcare facility that cannot let employee voices leave its premises. The Mistral API also provides simple access via the Python SDK, with voice management, streaming, and multiple audio formats, letting you prototype in a few lines before considering self-hosting. Finally, multilingual support is native: unlike solutions that add languages as an overlay, Voxtral was trained for many languages—35 with Voxtral TTS 3—with natural transitions between them—an English proper name in a French sentence no longer triggers the tonal break characteristic of older engines.

What you’ll learn

This course guides you through all of Voxtral TTS’s features. You’ll start by understanding the architecture and available models, then learn to create and manage custom voices via the API. Next, you’ll generate speech in real time with streaming before tackling latency optimization and choosing the right audio format. The journey concludes with building a complete voice pipeline, from speech recognition to synthesis via the LLM (STT → LLM → TTS).

As for prerequisites, you’ll need Python 3.10+ (the minimum required by the Mistral SDK since version 2.0) installed on your machine, a Mistral AI account with an API key, and the Mistral Python SDK obtained via pip install mistralai. Basic Python knowledge—variables, functions, libraries—is enough: no code in this course exceeds thirty lines.

Key takeaways

• Voxtral TTS is Mistral AI’s text-to-speech engine, launched in March 2026
• Zero-shot cloning reproduces a voice from an MP3 or WAV recording lasting a few seconds to a few minutes
• Voxtral TTS 3 covers 35 languages and more than 50 dialects, with approximately 90 ms before the first sound
• The previous generation’s weights are published under the CC BY-NC 4.0 license (non-commercial use)
• The Mistral API provides a complete Python SDK for voice management and speech generation