Mistral Voxtral 3 Brings a Voice Agent SDK to 50+ Languages

Mistral AI has launched Voxtral 3, a new family of speech models plus a voice agent SDK that ties listening, reasoning, tool calling and speaking into one real-time pipeline. Announced on October 20, 2026, the release targets the hardest part of voice AI: not just transcribing or speaking well, but holding a live conversation that actually gets work done, like checking a bill or booking an appointment mid-call.

• Transcription: Voxtral Small Transcribe 3 and the 4B-parameter Voxtral Mini Transcribe 3 cover 50+ languages, including all 24 official EU languages
• Realtime: Voxtral Mini Transcribe Realtime 3 targets sub-500ms latency in 35+ languages, with diarization and end-of-turn detection
• Speech output: Voxtral Speech 3 handles 35+ languages and 50+ dialects at $13 per million characters
• Pricing: $0.0035/min (Mini batch), $0.004/min (Small) and $0.008/min (Realtime) on Mistral's API

What Is New in Voxtral 3 Transcription?

The speech-to-text side gets three upgrades that matter in production. Diarization labels who said what, which is essential for meetings and contact-center calls. Context biasing lets developers pass in key terms, product names or speaker names so the model recalls domain vocabulary without fine-tuning. On Mistral's keyphrase test, supplying that context lifted Voxtral Small Transcribe 3's recall from 81.4% to 94.6%. And native end-of-turn detection uses both acoustic and semantic cues to judge when a caller has finished speaking, so an agent can respond quickly without cutting people off.

Mistral says Voxtral Small Transcribe 3 is state of the art in multilingual transcription and outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe on its FLEURS-based tests. Those comparisons come from Mistral's own benchmarking, so treat them as the company's claims until independent leaderboards catch up.

Expressive Speech and Pro Voice

The text-to-speech model reads the context of what it is saying and adjusts emotion, pacing and emphasis accordingly, and developers can also steer those variables with explicit instructions. A pronunciation hint system handles unusual names and brand terms. For businesses that want a distinctive branded voice, Mistral's Pro Voice adaptation reproduces a speaker's character and speaking style across languages from about 30 minutes of audio. That is a step beyond the open-weight Voxtral TTS model Mistral shipped earlier this year.

How Does the Voxtral Voice Agent SDK Work?

The SDK is built on LiveKit and runs speech in, reasoning and speech out as one streaming pipeline. The realtime transcriber listens, a general model such as Mistral Small reasons and calls tools, and Voxtral Speech 3 replies in the tone of the conversation so far. Because no stage waits for the previous one to finish, Mistral says each turn completes in under a second, and the agent stops talking when the caller interrupts.

Practical features include:

• Telephony: answering and placing calls over SIP, plus browser and app audio over WebRTC
• Actions: function calling, web search and retrieval over a company's own systems during a live call
• Swappable parts: each stage is a separate model, so teams can change the LLM, voice or language without rewriting the agent

It lands two days after xAI's Grok Voice 3, which shows how fast voice agents are becoming a core battleground for model makers.

Where You Can Run Voxtral 3

Voxtral 3 is available today through Mistral's API on European, US and global endpoints, and developers can try voice agents, transcription, speech generation and voice cloning in Mistral Studio's playgrounds. Organizations that need audio to stay in-house can talk to Mistral about running the full stack on-premises or in a private cloud, which fits the company's sovereign-deployment pitch for healthcare, finance and public-sector teams. For the latest model launches, see our AI coverage.

Sources: Mistral AI — Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages — October 20, 2026; Crypto Briefing — Mistral unveils new voice agent API and speech models — October 20, 2026.