Blog
Mistral launches Voxtral 3 speech models and a kit for building voice agents
Mistral AI released Voxtral 3 on 20 October 2026: transcription models covering 50+ languages, an expressive text-to-speech model and a LiveKit-based voice agent SDK that listens, reasons, calls tools and speaks in one pipeline. Prices start at $0.0035 per minute. Mistral's benchmark claims are its own and no Nigerian languages are listed.
French AI company Mistral AI has released Voxtral 3, a new family of speech models together with a software kit for building voice agents: AI systems that can hold a phone or web conversation, look things up and take actions while they talk. Mistral announced it on 20 October 2026 and says it is available now through its API.
What is in Voxtral 3
The release has three parts, according to Mistral's announcement:
- Speech-to-text (transcription). Two models, Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3, turn recorded audio into text in more than 50 languages. Mistral says the list includes all 24 official EU languages and many Asian and Indian languages. The models can label who said what, accept a list of special words (such as product names) to improve accuracy, and detect when a speaker has finished talking.
- Live transcription. Voxtral Mini Transcribe Realtime 3 is built for live calls. Mistral says it returns text in under half a second.
- Text-to-speech. A speech model, listed on the price table as Voxtral Speech 3, reads replies aloud in more than 35 languages and 50 dialects. Mistral says it changes emotion, pace and emphasis to suit the conversation. A "Pro Voice" option can copy a particular speaker's style from about 30 minutes of recorded audio.
The voice agent kit
The new voice agent SDK (software development kit) joins these parts into one live pipeline. It is built on LiveKit, an open-source system for real-time audio. One model listens, a general text model such as Mistral Small works out the answer and calls any needed tools, and the speech model replies. Mistral says the stages run at the same time instead of waiting for each other, so each turn finishes in under a second, and the agent stops speaking if the caller interrupts.
In Mistral's examples, an agent could confirm a caller's identity, check a bill or book an appointment, using web search, the company's own documents or other business software. It can take and make phone calls over SIP, the standard behind most internet phone systems, and work in browsers and apps over WebRTC. Mistral's documentation says developers can build with it in Python or TypeScript.
Price and where it runs
Mistral lists these prices in US dollars:
- Mini Transcribe 3 (batch): $0.0035 per minute
- Small Transcribe 3: $0.004 per minute
- Realtime transcription: $0.008 per minute
- Speech 3: $13 per million characters
- Voice agent API: $0.04 per minute with Mistral Small 4, or $0.08 per minute with other reasoning models
The models are served from European, US and global API endpoints. Organisations can also run the full stack on their own servers by arrangement with Mistral, which the company presents as a way to keep sensitive calls and customer data under their own control.
Claims to treat with care
Mistral says its Small transcription model beats rivals including OpenAI's Whisper, Scribe v2 and Google's Gemini 3.5 Transcribe, and that its speech model costs less than several competitors. Many of the comparison charts are Mistral's own tests, some scored by another AI model acting as judge. They have not been independently confirmed, and businesses should test with their own recordings, accents and phone lines.
For Nigerian users, language coverage is worth checking. The 45-language list Mistral used for its main transcription benchmark does not include Hausa, Yoruba, Igbo or other Nigerian languages, and its 18 preset voice languages do not include any sub-Saharan African language.
Why it matters
Call centres, banks, hospitals and government help lines are among the most likely early users of voice AI. Until now, building a good voice agent usually meant joining separate tools from different vendors. Mistral is offering the whole chain from one supplier, with the option to self-host, which competes directly with offerings from OpenAI, Google and ElevenLabs.