Mistral Voxtral 3: speech models and a voice agent SDK

Voxtral 3 is Mistral's new family of speech models plus an SDK for building voice agents, released on October 20, 2026. It covers batch transcription, streaming transcription with end-of-turn detection, and text-to-speech, and the SDK chains them with a language model and tool calls into one streaming pipeline. You can use it through Mistral's EU, US or global API endpoints, or run it on your own infrastructure through Mistral's team.

The interesting part for developers isn't any single model. It's that the whole loop a voice agent needs (listen, decide, act, answer) now comes from one vendor with one bill, while each stage stays swappable. Here is what's in the release, how the pipeline works, and what to check before you build on it.

What is Voxtral 3?

Voxtral 3 is the third generation of Mistral's audio models. A voice agent is a program that holds a spoken conversation and can take actions during it, like looking up a record or booking a slot. Building one usually means stitching together three separate services: speech-to-text, an LLM, and text-to-speech. Voxtral 3 ships all three pieces and the glue.

Diarization means labeling who spoke when, so a transcript of a meeting comes back split by speaker instead of as one block of text. The batch models also support context biasing: you pass a list of key terms or free-form context, and the model becomes more likely to spell product names, drug names or ticker symbols correctly, without any fine-tuning.

Language coverage grew a lot. Batch transcription goes from 13 languages in the previous generation to 50, including all 24 official EU languages and a number of Asian and Indic languages. Speech generation goes from nine languages to more than 35. The realtime transcription model covers fewer languages than the batch ones, so check the model page for your exact list.

How does the Voxtral voice agent pipeline work?

The SDK runs three stages as one stream:

1. Listen. Voxtral Mini Transcribe Realtime 3 turns incoming audio into text as it arrives and decides when the user has finished their turn.
2. Reason. A general LLM, such as Mistral Small 4, reads the transcript, decides what to do and calls tools: function calls into your backend, web search, or retrieval over your documents.
3. Speak. Voxtral TTS generates the reply as audio, in a tone that matches the conversation so far.

The stages don't wait for each other to finish completely. Speech can start playing while the rest of the reply is still being produced, which is how Mistral gets each turn under one second. If the caller starts talking while the agent is speaking, the agent stops. That behavior is called barge-in, and it's one of the things that makes a voice agent feel like a conversation instead of a phone menu.

Two transports are supported. WebRTC handles browsers and apps. SIP, the standard protocol behind phone calls, connects the agent to existing telephony, IVR and contact-center systems; at launch, phone integration is aimed at enterprise customers. Because the SDK sits on LiveKit, an open-source framework for real-time audio and video, the media handling is a known quantity rather than something new to learn.

The design choice worth noticing is that each stage is a separate model. You can change the LLM, the voice or the language without rewriting the agent. If your bottleneck is the reasoning step, you can swap in a faster model there; a diffusion LLM like Mercury 3 is one example of a model built specifically for that kind of latency budget.

Why does end-of-turn detection matter?

The hardest part of a voice agent is not transcription accuracy. It's knowing when to talk. If the agent waits for a fixed stretch of silence, it either interrupts people who pause to think or leaves awkward gaps after every sentence.

End-of-turn detection is the model's guess that the speaker is done. Voxtral's realtime model uses both acoustic cues (pitch, trailing off) and semantic cues (whether the sentence sounds complete) to make that call. "My account number is..." followed by a pause should not trigger a reply. "That's all, thanks" should, even if the pause is short.

One limitation to plan around: in a live session, speaker labeling and end-of-speech detection can't run at the same time. For a one-to-one agent that's fine. For a call with several people on one line, you'll have to pick.

What changed in text-to-speech?

Voxtral Mini TTS 3 runs in batch mode for things like narration and in realtime mode for conversations. It adds explicit controls for emotion, pacing and emphasis, phoneme hints for words it would otherwise mispronounce, and a vocabulary library so a company name is said the same way every time. Free-form style instructions in plain language aren't available at launch, so you work through those explicit controls instead.

There's also a Pro Voice option that builds a custom voice from about 30 minutes of recorded audio and keeps that speaker's style when it speaks other languages. Treat this the way you'd treat any voice cloning feature: use recordings you have clear permission to use.

When should you use Voxtral 3?

It fits well when:

• You need data residency. You choose where audio is processed and whether any of it is kept, and EU-hosted endpoints are a first-class option.
• You serve many languages, especially European or Indic ones that older speech models handled poorly.
• You want one integration for a phone or in-app agent instead of three vendors and your own orchestration.
• You transcribe domain-heavy audio like earnings calls or medical dictation, where context biasing pays off.

Things to check first:

• Timestamps. The batch models return segment-level timestamps, not word-level ones, and live transcripts don't carry explicit start and end times. If you build subtitles or click-to-seek transcripts, confirm this works for you.
• Benchmarks. Accuracy and price comparisons on the launch page come from Mistral's own evaluations. Run your own audio through it, with your accents and your background noise.
• Cost at volume. Per-minute pricing is easy to estimate before you commit:

# Rough monthly cost for a voice agent at Voxtral 3 list prices
calls_per_day = 2_000
avg_minutes = 3.5
days = 30
minutes = calls_per_day * avg_minutes * days

for label, per_min in [("SDK + Mistral Small 4", 0.04), ("SDK + larger model", 0.08)]:
    print(f"{label}: {minutes:,.0f} min -> ${minutes * per_min:,.2f}")

# SDK + Mistral Small 4: 210,000 min -> $8,400.00
# SDK + larger model: 210,000 min -> $16,800.00

The choice of reasoning model doubles the bill, so it's worth testing whether the smaller model handles your conversations before defaulting to the larger one.

A practical way to start: try the models in Mistral Studio's playground with recordings of real calls, then build a narrow agent with one or two tools (a lookup and a booking, say) over WebRTC before touching telephony.

Key takeaways

• Voxtral 3 bundles batch transcription, streaming transcription, text-to-speech and a voice agent SDK built on LiveKit.
• Batch transcription now covers 50 languages, and speech generation covers more than 35.
• The SDK streams listen, reason and speak stages together, supports barge-in, and targets under one second per turn.
• Every stage is a separate model, so you can swap the LLM, voice or language independently.
• Check timestamp granularity, the diarization and end-of-turn trade-off, and your own accuracy numbers before production.

FAQ

How much does Voxtral 3 cost?

Batch transcription starts at $0.0035 per minute, realtime transcription is $0.008 per minute, and text-to-speech is $13 per million characters. The full voice agent SDK is $0.04 per minute with Mistral Small 4, or $0.08 per minute with a larger model.

Can you self-host Voxtral 3?

Yes. Mistral offers Voxtral 3 on your own infrastructure, including private cloud, alongside its hosted EU, US and global API endpoints. The full on-premises stack is arranged through Mistral's sales team.

Does the Voxtral voice agent work over phone lines?

Yes, through SIP, which connects it to existing telephony and contact-center software. At launch, phone integration is aimed at enterprise customers, while WebRTC covers browsers and apps.