The AI Tool Shelf- Mistral launched Voxtral 3 on October 20, 2026: two transcription models, a text-to-speech model and a voice agent SDK.
- Transcription covers 50+ languages; text-to-speech covers 35+ languages and 50+ dialects, with voice cloning.
- Prices start at $0.0035 per minute for batch transcription and $13 per million characters for speech.
- It is a developer tool via API or self-hosting, not a point-and-click voiceover app like ElevenLabs or Murf.
Mistral AI, the French AI company, has launched Voxtral 3: a family of speech models for transcription and text-to-speech, plus an SDK for building real-time voice agents that can talk on the phone or in an app. It covers dozens of languages and comes with aggressive prices. Here is what Voxtral 3 includes, what it costs and how it compares to the voice tools creators already use. This is an explainer based on Mistral's announcement and press coverage.
What is in Voxtral 3?
Announced on October 20, 2026, the suite has three models:
- Voxtral Small Transcribe 3: speech-to-text, the most accurate of the family
- Voxtral Mini Transcribe 3: a smaller 4-billion-parameter speech-to-text model, with a real-time version
- Voxtral Mini TTS 3.0: text-to-speech
Languages and features
- Transcription in 50+ languages, with speaker identification (who said what), end-of-turn detection and support for custom key terms
- Text-to-speech in 35+ languages and 50+ dialects, with control over emotion, pace and emphasis
- Voice cloning: Mistral's Pro Voice option uses about 30 minutes of audio to reproduce a speaker's character and style
- Real-time speed: the real-time transcription model runs under 500 milliseconds, and voice-agent turns stay under one second
On accuracy, Mistral's own 45-language benchmark puts Small Transcribe 3 at about a 4% error rate, ahead of several well-known rivals, and an independent leaderboard (Artificial Analysis) ranked it third for speech-to-text at launch.
How much does it cost?
| Service | Price |
|---|---|
| Mini Transcribe 3 (batch) | $0.0035 per minute |
| Mini Transcribe 3 (real-time) | $0.008 per minute |
| Text-to-speech | $13 per million characters |
That makes transcribing an hour of audio cost around 21 cents in batch mode. These are API prices for developers.
The voice agent SDK
The SDK is built on LiveKit and lets developers assemble a phone or app assistant that listens, thinks, uses tools and speaks at the same time. It can stop talking when the caller interrupts, connect over regular phone lines (SIP) or the web (WebRTC), and call tools such as identity checks, appointment booking or web search. Each part is modular, so a business can swap the language, voice or reasoning model without rebuilding the agent. Companies can use Mistral's EU, US or global API, or run it in their own cloud to keep audio in-house.
Voxtral 3 vs ElevenLabs and Murf
Voxtral 3 is aimed at developers and companies: you use it through an API or Mistral's Studio playground. Creator tools like ElevenLabs and Murf AI give you a ready-made editor where you paste a script and download a voiceover, as in our ElevenLabs vs Murf test. ElevenLabs also offers its own voice agents for businesses, which we tested in our AI phone assistant guide.
So for a YouTube voiceover today, a creator tool is still the easier path. Voxtral 3 matters more if you are building an app, a call-centre assistant or need cheap transcription at scale, especially in Europe, where keeping data in the EU can be a requirement.
The bottom line
Voxtral 3 gives developers accurate transcription in 50+ languages, expressive text-to-speech, voice cloning and a kit for real-time voice agents, at low prices and with EU or self-hosted options. It is not a point-and-click tool for creators, but it adds strong competition to voice AI, which usually means better features and lower prices for everyone.
Sources: TestingCatalog, Technobezz, dutchstartup.ai.