# Voxtral > Voxtral: Mistral AI's audio model family (transcription, speech understanding and text-to-speech) used to talk to AI agents and hear them answer out loud. - **Category**: AI - **URL**: https://www.romaindelfosse.fr/en/glossary/voxtral/ --- **Voxtral** is Mistral AI's audio model family: it covers **transcription** (speech-to-text), **audio understanding** (Q&A, summarization, translation) and **speech synthesis** (text-to-speech). Launched in July 2025 with two open-weights speech-understanding models — Voxtral Small 24B for production and Voxtral Mini 3B for local and edge deployments, under the Apache 2.0 license — the lineup grew in March 2026 with **Voxtral TTS**, a 4-billion-parameter speech generation model. ## Technical definition ### Three use cases, one family **Transcription (Voxtral Mini Transcribe).** The API entry point for converting speech to text. With a 32k token context, the models ingest up to **30 minutes of audio for transcription** (40 minutes for understanding) in a single request, with automatic language detection across the world's most spoken languages — French, English, Spanish, Portuguese, Hindi, German, Dutch, Italian, Arabic. A feature particularly useful for agents: **function calling straight from voice** — a spoken intent can directly trigger a tool or API call, with no intermediate parsing step. **Audio understanding.** Beyond transcription, Voxtral answers questions about a recording's content or produces structured summaries — without chaining a separate ASR and then a language model. The model also retains the text capabilities of its backbone (Mistral Small 3.1). **Speech synthesis (Voxtral TTS, March 2026).** A 4B-parameter autoregressive flow-matching model (3.4B transformer decoder + 390M flow-matching acoustic transformer + 300M neural audio codec), built on Ministral 3B. It generates expressive speech in 9 languages, with a **70 ms model latency** (10 s sample, 500 characters), **voice adaptation from a 3-second reference**, and zero-shot cross-lingual adaptation — generating English with a French voice, for instance. Preset voices ship in American, British and French dialects (including *Marie* and *Jane*, the ones Mistral Vibe uses). ### Why it matters for agents A coding [AI agent](/en/glossary/agent-ia/) like [Mistral Vibe](/en/blog/m5stack-vibe-talkie-walkie-voxtral/) can thus close the voice loop: dictation transcribed by Voxtral on input, answer read out loud by the TTS on output. Keyboard and screen stop being the agent's only interfaces — which is the precondition for remote control, away from the desk. And because the weights are open (Apache 2.0 on the understanding side), the audio brick stays auditable and deployable on your own infrastructure — a rare [digital sovereignty](/en/glossary/souverainete-numerique/) point in the voice market, dominated by closed proprietary APIs. ## Key Figures - Voxtral TTS model latency for a 10-second voice sample and 500 characters, with a real-time factor (RTF) of about 9.7x *(Mistral AI — Speaking of Voxtral (Voxtral TTS announcement) — March 23, 2026)* - Voxtral API transcription pricing per minute of audio, with Voxtral Mini Transcribe outperforming OpenAI Whisper for less than half the price *(Mistral AI — Voxtral (announcement) — July 15, 2025)* - Average listener preference for Voxtral TTS over ElevenLabs in a zero-shot custom-voice side-by-side test *(Mistral AI — Speaking of Voxtral (Voxtral TTS announcement) — March 23, 2026)* ## Frequently Asked Questions ### What is Voxtral? Voxtral is Mistral AI's audio model family. Launched in July 2025 on the speech-understanding side (Voxtral Small 24B and Voxtral Mini 3B, Apache 2.0 license), it covers transcription (speech-to-text), audio understanding (Q&A and summarization directly over a recording) and, since March 2026, speech synthesis (Voxtral TTS, a 4-billion-parameter model). The models natively handle French, English and seven other languages. ### What is the difference between Voxtral Transcribe and Voxtral TTS? Voxtral Transcribe converts speech to text: up to 30 minutes of audio per request, with automatic language detection and voice-triggered function calling. Voxtral TTS does the opposite: it generates realistic, expressive speech from text, with voice adaptation from as little as a 3-second reference and 70 ms model latency. Combined, the two enable end-to-end voice agents. ### How does Mistral Vibe use Voxtral? The Mistral Vibe coding agent relies on Voxtral for its voice loop: push-to-talk dictation is transcribed by Voxtral, and the agent's final answer can be read out loud by the embedded TTS client (voxtral-mini-tts-latest model, Marie voice in French and Jane in English). That's the mechanism the vibe-m5stack project uses to turn an M5Stack Fire into an agent walkie-talkie. ## Resources - [Voxtral — speech understanding model announcement](https://mistral.ai/news/voxtral) — Mistral AI - [Speaking of Voxtral — Voxtral TTS announcement](https://mistral.ai/news/voxtral-tts/) — Mistral AI - [Voxtral — research paper](https://arxiv.org/abs/2507.13264) — Mistral AI (arXiv)