Voxtral in 2026: TTS and Real-Time Transcription

Updated November 2, 2026 — audio. The Voxtral TTS (voxtral-mini-tts-2603), Voxtral Mini Transcribe 2 (voxtral-mini-2602), and its real-time version models have been deprecated since October 31, 2026, with removal scheduled for April 30, 2027. Their successors in the Voxtral 3 family, released October 20 in public preview, and the Voice Agents pipeline are covered in the Mistral in 2026 course (lesson 11).

Mistral’s audio capabilities expanded in 2026 with two new building blocks: text-to-speech synthesis and real-time transcription. Taken individually, these are just two technical announcements among many; taken together, they complete a chain—from the user’s voice to the machine’s voice—that now lets you build a complete voice agent without leaving the Mistral ecosystem. This lesson covers each component, then shows how they fit together.

Voxtral TTS: Giving Your Applications a Voice

Voxtral TTS (v26.03), announced March 23, 2026, is Mistral’s text-to-speech synthesis model. The official documentation highlights state-of-the-art quality for voice agents, multilingual support, and above all, a capability worth noting: zero-shot voice cloning. Specifically, the model can reproduce a voice from a short audio sample without dedicated training. Where voice cloning previously required hours of recording and a specific learning phase, a few seconds now suffice to obtain a convincing imitation.

This ease of use demands an immediate reminder, and this is not a formality: cloning a person’s voice requires their explicit consent. In Europe, voice is personal data under the law. Impersonation uses—making someone say what they never said, posing as them on the phone—are illegal, period. As for legitimate uses, and there are very solid ones—a consistent brand voice across all channels, accessibility solutions that restore voice to people who have lost theirs—they must be governed by contract: written consent, defined scope of use, specified duration. If your project involves cloning a real voice, build this legal framework before writing the first line of code, not after.

Voxtral Mini Transcribe Realtime: Understanding in Real Time

The second component addresses a clearly identified gap. Mistral already offered audio file transcription with Voxtral, but this batch mode assumes the recording is finished before processing. However, a whole family of use cases cannot wait: live meeting or event captioning, a voice agent that must understand while the user is speaking, continuous automatic note-taking.

This is what Voxtral Mini Transcribe Realtime (v26.02) covers, described in the documentation as an efficient input audio model, pre-trained and optimized for live transcription. The difference from batch mode is not cosmetic: in real time, latency becomes the primary constraint, because a voice agent that takes three seconds to understand a sentence produces an unbearable conversation. A model specifically optimized for the continuous stream is therefore the right architectural response, rather than repurposing the batch model.

The Mistral Audio Suite, and How to Assemble It

The audio lineup now consists of three complementary components:

For a complete voice agent, the typical chain assembles as follows: real-time transcription converts the user’s speech to text on the fly, a language model—Small 4 for most cases, Medium 3.5 for the most demanding—understands and formulates the response, then Voxtral TTS speaks it. Every link comes from the same provider, which simplifies contracting, billing, and compliance—an argument that carries weight in enterprise projects.

Our “Voxtral Transcription” and “Voxtral TTS” courses cover the fundamentals of these components; this lesson places them in the updated 2026 range. If you have taken them, the architecture described here is the natural combination of what you already know how to do component by component.

Official source: docs.mistral.ai — Models Overview

Key Takeaways

• Voxtral TTS (v26.03): State-of-the-art synthesis with zero-shot voice cloning
• Voice cloning requires the person’s explicit consent—this is non-negotiable
• Voxtral Mini Transcribe Realtime (v26.02) covers live transcription
• Voice agent chain: real-time transcription → LLM → TTS, all from the same provider