Mistral Ships Voxtral 3: Voice Agent SDK with Sub-Second Turns Across 50 Languages
Mistral released Voxtral 3 on October 20, a five-model family covering the full voice pipeline from transcription to speech synthesis, plus a voice agent SDK that wires all three stages into a single streaming loop.
The practical result: a phone agent that listens, reasons, calls tools, and speaks — completing each turn in under one second end to end. Mistral says the SDK handles the transport and latency management so developers do not need to build their own realtime audio pipeline.
Five Models, Three Jobs
Voxtral 3 ships five distinct models:
- Voxtral Mini Transcribe 3 — batch STT, speaker diarization, segment timestamps, custom vocabulary. 65,536-token context shared between audio and vocabulary list.
- Voxtral Mini Transcribe Realtime 3 — streaming STT for live audio, with configurable turn detection. Latency ranges from 80ms to 2,400ms per session setting.
- Voxtral Small Transcribe 3 — higher-accuracy batch transcription; transcription only at launch.
- Voxtral TTS 3 — text-to-speech with preset or cloned voices, 34 emotion tags, 2 pace tags, WAV at 24 kHz output.
- Voxtral TTS Realtime 3 — streaming speech generation for live conversations.
Transcription covers 50+ languages. TTS supports 35 languages and 50+ dialects with voice presets in 18, including Arabic (Emirati and Modern Standard), Mandarin, Hindi, Korean, and major European languages.
SDK Architecture
The voice agent SDK builds on LiveKit for realtime audio transport. Mistral adds Python and TypeScript bindings and the bridge to its models. The pipeline runs three stages in parallel: Voxtral Mini Transcribe Realtime 3 listens and detects turn ends, Mistral Small (or another choice of reasoning model) handles inference and tool calling, and Voxtral TTS Realtime 3 streams the reply.
Agents can call tools during a live call — checking a bill, verifying identity, booking an appointment — over function calling, web search, or retrieval against a private data store. Callers can interrupt the agent mid-sentence. SIP connectivity for inbound and outbound calls is available to enterprise accounts; browser and app use runs over WebRTC.
Because the SDK separates each stage into an independent model, swapping the LLM, the voice, or the transcription language does not require rewriting the agent. That composability is relevant for multilingual contact-center deployments where the same pipeline needs to handle different caller languages in different regional instances.
Pro Voice
For professional use cases, Mistral adds Pro Voice adaptation: voice cloning that reproduces a speaker’s emotional range, pace, and accent from 30 minutes of source audio. Unlike basic voice cloning that captures tone, Pro Voice is designed to stay consistent across languages — the same synthesized voice speaking in French and German, not two separate voice personas.
Pricing
| Model | Price |
|---|---|
| Voxtral Mini Transcribe 3 (batch) | $0.0035/min |
| Voxtral Mini Transcribe Realtime 3 | $0.008/min |
| Voxtral Small Transcribe 3 | $0.0040/min |
| Voxtral Mini TTS 3 | $13/1M chars |
The realtime transcription rate of $0.008/min places Voxtral 3 in the same bracket as Deepgram and AssemblyAI’s streaming products. At $0.0035/min for batch, it undercuts OpenAI Whisper-based competitors on the low end of production workloads.
Data Sovereignty
Voxtral 3 is available through Mistral’s EU-hosted API and through US and global endpoints, with self-deployment as an alternative. Enterprise customers can route audio entirely on-premises, which matters for healthcare, legal, and financial services deployments where call content cannot leave a controlled environment. Mistral’s European HQ gives it regulatory standing in GDPR-sensitive enterprise sales that US-only voice AI providers cannot match without additional contractual overhead.
Voice agent builds on large language models have so far required teams to stitch together separate vendors for STT, LLM, and TTS. Voxtral 3’s single-vendor stack reduces integration surface and gives Mistral a competitive entry into the contact-center automation market where per-minute billing on unified pipelines is the standard purchasing model.