Voxtral — Mistral's Audio Models

Audio, the New Frontier of AI

Artificial intelligence is no longer limited to text. Understanding and producing speech has become a concrete challenge for many applications: voice assistants, meeting transcription, dubbing, accessibility. Mistral developed the Voxtral family to cover this entire chain. The four models we’ll look at are not interchangeable: each occupies a specific point in the chain, and confusing their roles is the most common mistake when designing an application.

Voxtral TTS — Next-Generation Speech Synthesis

Launched in March 2026, Voxtral TTS is the text-to-speech synthesis model. It’s one of the year’s most notable releases, mainly due to one capability: zero-shot voice cloning.

The principle is simple to state and carries significant implications. You provide an audio sample of a voice—a few seconds are enough—and the model then generates any text in that voice, without additional training. The timbre, rhythm, and characteristics of the source voice are captured in a single pass: there is no need to fund data collection or training sessions between hearing a voice and being able to use it.

Four application areas immediately emerge. Dubbing and localization allow you to translate audio content while preserving the original voice, maintaining brand or speaker identity across languages. Accessibility benefits from interfaces equipped with a natural, personalized voice rather than impersonal synthesis. Educational content finds consistent narration from one module to another, without re-recording. And voice assistants can finally have a distinctive sound identity.

Technically, the model is multilingual and supports French. It works in real-time streaming: the voice is produced on the fly, without waiting for the full text to be available, enabling fluid conversation. The output is close to human speech, with intonation and emotion management. The model belongs to the Open tier, with weights published on Hugging Face.

This power demands discipline. Mistral recommends systematically obtaining consent from the person whose voice is cloned, never using this technology to impersonate an identity, and clearly identifying AI-generated content. These three rules are not theoretical precautions: a dubbing project that neglects them exposes its client just as much as it exposes the person whose voice served as the model.

Voxtral Mini Transcribe 2 — Audio Transcription

Released in February 2026, Voxtral Mini Transcribe 2 goes the opposite way: it is Mistral’s audio transcription model, in other words speech-to-text, and it converts speech to text with high accuracy. It falls under the Premier tier, so it is accessed via the API.

Four capabilities distinguish it from a basic transcriber. Transcription is multilingual with automatic language detection, which avoids configuring each file in advance. Diarization identifies different speakers in a conversation—without it, a meeting report with six participants remains a block of unusable text. Context biasing allows you to provide the model with specialized vocabulary, proper names, or technical terms, so it stops mangling the references that matter in your field. Timestamps, finally, align the transcription word-by-word with the audio track.

These characteristics outline the use cases: meeting reports with participant identification, synchronized subtitling, customer support call analysis, and indexing spoken content to make a previously unsearchable archive of recordings searchable.

Voxtral Mini Transcribe Realtime — Real-Time Processing

When latency takes precedence over finesse, Voxtral Mini Transcribe Realtime transcribes in a continuous stream: text appears as the person speaks, without waiting for the end of the sentence. This model falls under the Open tier, which allows local hosting—a decisive argument for sensitive applications where audio must not leave your infrastructure.

Read this table as a trade-off, not as a ranking: you don’t choose Realtime because it’s better, but because live subtitling tolerates no post-processing.

Voxtral Small — Audio Input

Voxtral Small, published in July 2025, plays a different role from the previous ones. It accepts audio as input for instruction tasks: you send it a file and ask it to summarize it, answer questions about its content, or analyze it. So it is not a transcriber that you would chain with a text model, but an audio multimodal instruct model, which works directly on the sound signal.

Key Takeaways

• Voxtral TTS (March 2026) is the speech synthesis model with zero-shot voice cloning
• Voxtral Mini Transcribe 2 offers accurate transcription with diarization and context biasing
• Voxtral Realtime enables streaming transcription, open source
• Voxtral Small is a multimodal instruct model that understands audio input
• The Voxtral family covers the entire chain: text → voice (TTS), voice → text (transcription), and voice → understanding (instruct audio)