AI Industry
Mistral Voxtral 3 Brings Voice Agents in 50+ Languages
Oct 21, 2026·4 min read·By Rahish Saifi
Mistral AI on October 20, 2026 launched Voxtral 3, a full voice stack for production agents: new speech-to-text and text-to-speech models, plus a voice agent SDK built on LiveKit that streams listening, reasoning, tool use, and spoken replies in one pipeline. The Paris-based company is pitching contact-center agents, service desks, virtual assistants, and hands-free copilots—use cases where latency, language coverage, and data residency matter as much as raw accuracy.
Key facts
| Item | Detail |
|---|---|
| Launch | October 20, 2026 — Mistral announcement and docs |
| Stack | Speech-to-text models, TTS models, and a voice agent SDK (LiveKit-based) for real-time agents |
| Transcription | 50+ languages; diarization, keyphrase context biasing, end-of-turn detection |
| Speech output | 35+ languages and 50+ dialects; emotion, pace, and emphasis from context |
| Agent turns | Mistral says each turn completes in under one second; agent stops when the caller interrupts |
| Deployment | Mistral API (European, US, global endpoints) or self-deployment; control over audio retention |
| Voice agent API | $0.04/min with Mistral Small 4 reasoning; $0.08/min with any other reasoning model |
What Mistral launched
Mistral frames Voxtral 3 around four simultaneous jobs: listen accurately, reason in context, speak naturally, and act through tools over multiple steps. Instead of one end-to-end speech model, the stack uses separate transcription, reasoning, and TTS services that stream together.
Like other Mistral releases, it emphasizes sovereign deployment: EU-hosted API endpoints, US and global routes, or self-hosting, with customer control over audio retention. Mistral Studio playgrounds cover transcription, speech, and agent demos before production SIP or WebRTC rollouts.
The five Voxtral 3 models
Mistral's documentation lists five model roles in the Voxtral 3 family. Pricing below comes from Mistral's announcement and API docs as of launch.
| Model | Job | Price (Mistral) |
|---|---|---|
| Voxtral Mini Transcribe 3 | Batch file transcription (4B parameters) | $0.0035/min |
| Voxtral Mini Transcribe Realtime 3 | Live transcription for agents (sub-500ms latency per announcement; configurable streaming delay in docs) | $0.008/min |
| Voxtral Small Transcribe 3 | Highest-accuracy multilingual transcription and diarization | $0.004/min |
Voxtral TTS 3 (Voxtral Speech 3; id voxtral-mini-tts-3-0) | Batch and realtime expressive speech | $13 per 1M characters |
| Voxtral TTS Realtime 3 | Low-latency spoken replies in agent pipelines | Same Speech 3 pricing tier |
Mini Transcribe 3 handles batch files; Small Transcribe 3 targets maximum accuracy; realtime models feed agents that need end-of-turn detection and partial transcripts.
How the voice agent SDK works
The voice agent SDK stitches separate models into one streaming experience. In Mistral's reference architecture, Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has finished speaking. A general reasoning model—Mistral cites Mistral Small as an example—interprets intent, calls tools, and plans the reply. Voxtral Speech 3 then speaks with tone matched to the conversation so far.
Stages overlap in the stream, so Mistral says turns finish in under one second and the agent stops when the caller interrupts. Function calling, search, and retrieval can verify identity, check bills, or book appointments. Calls run over SIP or WebRTC; each stage can be swapped independently. SDKs ship for Python and TypeScript/Node.js; enterprise SIP goes through Mistral's account team at launch.
Languages, voices, and the India angle
Transcription now covers 50+ languages, including all 24 official EU languages and many low-resource Asian and Indic languages, per Mistral. TTS supports 35+ languages and 50+ dialects, with emotion, pace, and emphasis inferred from context.
Voice presets are available in 18 languages, including Hindi (alongside major European and Asian locales). Pro Voice adaptation, using about 30 minutes of audio, targets speakers whose style must carry across languages—beyond quick voice cloning.
Pricing
Transcription runs from $0.0035/min (Mini batch) to $0.008/min (Mini realtime) and $0.004/min (Small). Speech is $13 per 1M characters for Voxtral Speech 3. The bundled voice agent API costs $0.04/min with Mistral Small 4 reasoning or $0.08/min with any other reasoning model.
How it compares
On its own benchmarks, Mistral says Small Transcribe 3 beats Whisper, Scribe v2, and Gemini 3.5 Transcribe; realtime Mini transcription targets sub-500ms latency; and $13/1M characters TTS costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash, per Mistral's announcement. Treat those as vendor claims until you test your audio. Architecturally, the LiveKit SDK sits alongside other realtime voice stacks such as xAI's Grok Voice 3.0, with different pricing and model choices.
What developers and businesses should know
Start in Mistral Studio, then wire the API or ask Mistral about self-deployment. Enterprise SIP requires the account team at launch; WebRTC suits browser pilots. Audio retention is configurable—relevant for EU sovereignty rules. Modular stages let you keep an existing LLM or swap voices without rebuilding the agent, similar to how teams mix other Mistral services such as Timestral forecasting models for non-voice workloads.
Bottom line
Voxtral 3 bundles transcription, expressive TTS, and a LiveKit agent SDK for contact centers and copilots that need low-latency, multilingual voice with deployment choice—API or on-prem—and Hindi among preset voices. Whether it wins your stack depends on real calls, but Mistral has put a full spoken pipeline on the table in one launch.