> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Mistral Releases Voxtral 3 With Multilingual Speech Models and LiveKit Voice Agent SDK - URL: https://www.tao.media/mistral-releases-voxtral-3-with-multilingual-speech-models-and-livekit-voice-agent-sdk/ - Published: 2026-10-21T13:32:35.000Z - Updated: 2026-10-21T13:32:36.000Z - Description: The new stack connects transcription, reasoning, and expressive speech for phone and web agents, with API and private deployment options. - Author: Bart Hillerich - Tags: Mistral, AI, News Mistral AI has [released Voxtral 3](https://x.com/MistralAI/status/2112574615318876178?ref=tao.media), a family of transcription and speech-generation models paired with a LiveKit-based software development kit for voice agents. The release connects speech recognition, a reasoning model, tool calls, and spoken responses in a streaming pipeline for phone and web conversations. The stack targets contact centers, service desks, virtual assistants, and hands-free copilots. Agents can retrieve information from business systems and take actions during a call, such as checking a bill or booking an appointment. Mistral says the pipeline completes conversational turns in under one second and stops speaking when a caller interrupts. [Voxtral 3](https://mistral.ai/news/voxtral-3?ref=tao.media) is available through Mistral's European, U.S., and global API endpoints, with self-deployment options for organizations that need to process audio on their own infrastructure. The release also introduces Pro Voice, a professional voice-adaptation service that reproduces a speaker's vocal character across languages. ## Transcription Adds Domain Vocabulary and Turn Detection The transcription lineup separates higher-accuracy processing, lower-cost batch work, and live audio. Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 support 50 languages, including all 24 official European Union languages and a range of low-resource Asian and Indic languages. The Mini batch model has 4 billion parameters. Both models support diarization, which identifies who spoke when, and context biasing. Developers can supply key terms or free-form background text to improve recognition of specialized vocabulary without fine-tuning the model. For a service desk, that could mean supplying product names or technical terms that would otherwise be difficult to distinguish from ordinary words. ![Voxtral Small Transcribe 3 transcription demo](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/Fvc45NNY8A1C5nKr.jpg) [Voxtral Small Transcribe 3 transcription demo | Mistral AI on X](https://x.com/MistralAI/status/2112574618586247583?ref=tao.media) Voxtral Mini Transcribe Realtime 3 handles incoming audio as a stream. Mistral reports sub-500-millisecond transcription latency in more than 35 languages. The [realtime model documentation](https://docs.mistral.ai/models/voxtral-mini-transcribe-realtime-3-0?ref=tao.media) lists it as a public preview and specifies configurable latency from 80 to 2,400 milliseconds, with diarization, context biasing, and end-of-turn detection available in a single WebSocket session. End-of-turn detection uses both acoustic and semantic information: how an utterance sounds and what it means. That lets an agent estimate when someone has finished a request rather than relying only on a pause, reducing the chance that it starts answering while the caller is still speaking. Mistral reports that Small Transcribe 3 outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe in its evaluations. Its multilingual comparison covers 45 shared languages from FLEURS, with separate results for a specified language and automatic language detection. It measures word error rates for most languages and character error rates for Japanese, Korean, Thai, and Chinese. The company also evaluates speaker-attributed transcription across public conversation datasets and keyphrase recall on Earnings-22, including tests with supplied vocabulary and distractor terms. Those tests address different operational needs: recognizing the words, assigning them to the right speaker, and preserving names or specialized terminology. ## Speech Generation Adds Expression and Pronunciation Controls Voxtral Mini TTS 3 generates speech in batch and realtime modes, supporting more than 35 languages and more than 50 dialects. It can infer emotion, pace, and emphasis from conversational context, while explicit instructions let developers set those qualities line by line. Pronunciation controls include phoneme hints and a vocabulary library. These give developers a way to specify how unusual names, branded terms, or ambiguous words should sound instead of changing the surrounding sentence to coax the desired pronunciation. Mistral also provides voice presets across multiple languages and regional accents. Mistral's speech evaluations include Voice Code Bench, developed by BeSimple AI to test customer-support failure cases involving numbers and unique names. The company's run uses 300 English prompts and two voices per system. Additional internal evaluations cover pronunciation, multilingual quality, and expressive English, using Gemini 3.8 Flash as the judge. Pro Voice extends the release into [professional voice adaptation](https://x.com/MistralAI/status/2112574624034677048?ref=tao.media). Using roughly 30 minutes of recorded speech, it aims to reproduce a speaker's distinctive character, emotional range, and speaking style across supported languages, rather than only matching a basic vocal identity. ## A Modular Voice Agent Pipeline The [LiveKit-based SDK](https://x.com/MistralAI/status/2112574627381936345?ref=tao.media) combines separate models rather than requiring a single model to handle the entire interaction. Realtime Transcribe listens and detects turn boundaries; a general-purpose model, such as Mistral Small 4, reasons and calls tools; and Voxtral TTS produces the response. ![Voxtral 3 voice agent pipeline diagram](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/diagram-voxtral-alt_Z2ufcLf.webp) [Voxtral 3 voice agent pipeline diagram | Mistral AI](https://mistral.ai/news/voxtral-3?ref=tao.media) The stages stream their outputs rather than waiting for each preceding stage to finish completely. Mistral attributes its under-one-second conversational-turn claim to that overlapping execution. The architecture also lets developers change the reasoning model, voice, or language without rewriting the agent. Function calling, web search, and retrieval connect the conversation to external information and business workflows. Mistral's demonstration shows a caller providing a corrected referral number, the agent retrieving a French medical referral, explaining it in English, and using a booking tool to confirm an appointment. For deployment, the SDK supports incoming and outgoing calls over SIP and browser or app sessions over WebRTC. Those protocols connect the agent to telephone systems and live web audio, respectively. Mistral says the stack can integrate with existing telephony, interactive voice response, and contact-center software. ## API Pricing and Private Deployment Mistral prices transcription by audio minute and speech generation by text volume. The voice agent API has separate per-minute rates depending on the reasoning model used. | Model or Service | Published API Price | | ------------------------------------ | ---------------------------- | | Voxtral Mini Transcribe 3, batch | $0.0035 per minute | | Voxtral Small Transcribe 3 | $0.0040 per minute | | Voxtral Mini Transcribe Realtime 3 | $0.008 per minute | | Voxtral Mini TTS 3 | $13 per 1 million characters | | Voice agent API with Mistral Small 4 | $0.04 per minute | | Voice agent API with a larger model | $0.08 per minute | Developers can test voice agents, transcription, speech generation, and voice cloning in Mistral's Studio playgrounds. For on-premises or private-cloud deployment of the full stack, Mistral directs organizations to its team. Self-deployment gives customers control over where conversations are processed, how audio is handled, and whether recordings are retained.