Mistral Voxtral 3 Voice Models and Agent SDK

Mistral shipped Voxtral 3 on October 20, 2026, per its official launch page stamped 12:00 UTC. Voxtral Mini Transcribe 3 batch audio lists at $0.0035 per minute.

Voxtral Mini Transcribe Realtime 3 lists at $0.008 per minute on the same price table.

LiveKit agent pipeline

The voice agent SDK sits on LiveKit and runs three stages as one stream. Voxtral Mini Transcribe Realtime 3 listens and fires end-of-turn detection from both acoustic and semantic cues. A reasoning model such as Mistral Small 4 decides what was meant and issues tool calls. Voxtral Speech 3 (also labeled Voxtral Mini TTS 3.0 in the same post) speaks the reply in the tone of the turn so far.

No stage waits for the prior one to finish. Mistral says each turn completes in under one second, and playback stops when the caller cuts in. The same stack answers and places calls over SIP, runs in browsers and apps over WebRTC, and plugs into existing telephony and IVR gear. Because each stage is a separate model, you can swap the LLM, the voice, or the language without rewriting the agent graph.

A voice agent API session billed with Mistral Small 4 for reasoning is $0.04 per minute. The same agent line with any other reasoning model is $0.08 per minute. Pro Voice adaptation, aimed at brand and professional cloning where a short sample is not enough, needs 30 minutes of reference audio. Context biasing on the transcribers takes keyterms and free-form text so domain vocabulary lands without a fine-tune pass.

During a live call the agent can verify a caller's identity, check a bill, or book an appointment through function calling, web search, and retrieval against your own systems.

Diarization cpWER chart

Mistral’s headline transcription claim is multilingual accuracy plus speaker attribution. On its diarization chart (macro-average concatenated minimum permutation WER over AMI-IHM, AMI-SDM, Switchboard, Switchboard CallHome, and AISHELL-4, using 30-minute AMI and AISHELL cuts for API baselines), Voxtral Small Transcribe 3 prints 18.4%, against 38.4% for Gemini 3.5 Transcribe, as CellCog’s read of the same charts records. On FLEURS across 45 overlapping languages, Small Transcribe shows 4.1% word error with the language forced. Mistral writes that the Small model outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe. CellCog notes the chart also ties Microsoft AI Transcribe 2 at that same forced-language mark, a detail the surrounding sentence skips.

Realtime transcription is sold as sub-500ms latency for agent work. CellCog’s extract of Mistral’s realtime FLEURS curve is the useful caveat: error falls as the model waits longer before committing text, and Gemini 3.5 Realtime Transcribe stays ahead at every latency setting on that plot. Speech quality charts are internal evals scored by gemini-3.8-flash on prompts Mistral selected for release, with prices taken from vendor pages and a generation stamp of 2026-10-20. Artificial Analysis’s WER Index appears as another weighted transcription view on the launch page, still inside Mistral’s presentation.

What's underappreciated here: gemini-3.8-flash is the judge on those TTS evals, and the prompts are ones Mistral selected for release.

On-prem sales path

The post aims the stack at contact centers, service desks, virtual assistants, and hands-free copilots that need tool use mid-call. Transcription coverage is set at 50 languages, including the 24 official EU languages. Speech generation covers 35+ languages and 50+ dialects, with voice presets listed for 18 languages. Voxtral Small Transcribe 3 is the accuracy tier at $0.0040 per minute. Voxtral Mini Transcribe Realtime 3 is the live tier at $0.008 per minute. Voxtral Speech 3, batch and realtime on the same API, is $13 per 1M characters. Mistral says that speech rate undercuts comparable lines from OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash on list price.

Access is live today through European, US, and global API endpoints, with Studio playgrounds for agents, transcription, speech, and cloning. Self-deployment on your own infrastructure is offered as a sovereign path. The on-prem and private-cloud route goes through Mistral’s sales team rather than a public weights download.

No independent lab has republished the 18.4 diarization figure on outside data, and the launch post does not publish weights for any Voxtral 3 checkpoint.