Voxtral 3: Mistral’s voice agent stack, explained

Voxtral 3 is in public preview, so prices, limits and model names may change.

The answer in brief

Voxtral 3 is Mistral AI’s third generation of speech models, released on 20 October 2026. It pairs three transcription models and a speech generation model with a voice agent SDK built on LiveKit, so one stack can run a whole conversation, from the caller’s first word to the spoken reply. Mistral’s changelog lists four models in public preview on the Mistral API, which has global, EU and US endpoints.

• Transcription is near the top of the field. As of 24 October 2026, Voxtral Small Transcribe 3 ranks third on Artificial Analysis’s speech-to-text leaderboard, with a 2.1% word error rate. Both batch models now handle 50 languages, up from 13.
• Speech is cheap but trails the leaders. Voxtral TTS 3 costs $13 per million characters, far below Eleven v4 and Cartesia Sonic 3.6, and ranks 12th of 98 in Artificial Analysis’s blind English arena.
• A hosted agent is listed at $0.04 a minute with Mistral Small 4. Mistral does not say what that rate includes. Transcription starts at $0.0035 a minute. These are Mistral’s US dollar list prices; its pricing page also shows euro prices and does not say whether tax is included.
• No open weights this time. Earlier Voxtral releases put models on Hugging Face. Voxtral 3 runs on your own servers only through Mistral’s sales team.
• Phone support is narrow. The agent answers inbound SIP calls only, cannot transfer a caller to a person, and SIP is open only to enterprise customers.

The release is aimed squarely at contact centres, service desks and voice assistants, especially in organisations that want audio processed in the EU. Voxtral 3 is a strong transcription choice today; test the speech and the phone line against your own calls before you commit to the rest. Mistral’s EU endpoint and zero data retention both exclude Agents, and neither page says where the new Voice Agents fit, so ask Mistral before a pilot.

What does Voxtral 3 include?

Mistral’s audio documentation lists five models and one SDK. The changelog and price list name only four: the realtime speech model has no entry of its own, so treat its availability and price as unconfirmed.

The voice agent SDK joins the realtime transcriber, a reasoning model and the speech model into one streaming pipeline. Sources: Mistral audio docs, languages page, announcement, checked 24 October 2026.

Mistral’s own pages call the speech model both Voxtral Mini TTS 3 and Voxtral TTS 3, and some coverage calls it Voxtral Speech 3. Mistral has not published a parameter count for Small Transcribe 3 or for the TTS models.

How accurate is Voxtral 3 transcription?

Voxtral Small Transcribe 3 ranks third on Artificial Analysis’s non-streaming speech-to-text leaderboard, with a 2.1% word error rate, behind Alibaba’s Fun-Realtime-ASR-preview (1.7%) and Microsoft AI Transcribe 2 (2.0%). Mini Transcribe 3 scores 2.2%, level with ElevenLabs Scribe v2. Mistral’s own charts, below, cover different tasks and more languages.

Both batch models handle 50 languages, up from 13 in Voxtral Transcribe 2, including all 24 official EU languages and several low-resource Asian and Indic languages. You can pin the language with a two-letter ISO code or let the model detect it.

Diarisation, labelling who spoke when, also improves. On Mistral’s speaker-attributed test, which uses cpWER, Small Transcribe 3 scores 18.4% against 30.9% for its predecessor, Voxtral Mini Transcribe 2. You can now also send up to eight reference clips of known voices, each 3 to 15 seconds long, so a transcript labels a speaker by name rather than as speaker_1. The total number of speakers, counting references and those the model finds, cannot exceed 10.

Source: Mistral’s Voxtral 3 announcement, 20 October 2026. Vendor-run benchmark, macro-averaged over AMI-IHM, AMI-SDM, Switchboard, CallHome and AISHELL-4, using 30-minute versions of AMI and AISHELL.

On Mistral’s 45-language FLEURS test, Small Transcribe 3 scores 4.0% to 4.1% depending on whether the language is specified or detected, level with Microsoft AI Transcribe 2, while Whisper large-v3 scores 14.9% to 16.0%. The average uses character error rate for Japanese, Korean, Thai and Chinese and word error rate for the rest. The Small Transcribe 3 model card goes further, claiming the model outperforms Microsoft AI Transcribe 2 as well as Whisper, GPT-Transcribe, Scribe v2 and Gemini 3.5 Transcribe. Mistral’s own FLEURS chart does not show a win over Microsoft: the two sit level.

Context biasing lets you pass product names, jargon or a free-form sentence so the model favours them, with no fine-tuning. Entries can be multi-word phrases. On Earnings-22 keyphrase recall, Small reaches 94.6% with context and 81.4% without. Scribe v2 is barely behind at 94.3% with context and does better without it, at 84.0%.

What about live transcription?

Voxtral Mini Transcribe Realtime 3 streams text over a WebSocket with a look-ahead delay you choose per session, from 80 ms to 1,200 ms, or 2,400 ms. Network and application delays come on top. Its end-of-turn detection weighs both the sound and the meaning of what was said, so an agent can reply quickly without cutting a caller off mid-thought. Realtime diarisation is new in this generation, with a catch: a single realtime session can label speakers or detect turns, not both, and the voice agent SDK does not offer diarisation at all. Mistral’s advice is to run turn detection live, keep the audio and diarise the recording afterwards. Realtime segments also carry no start or end times.

On Mistral’s 35-language realtime FLEURS chart, the model scores 6.4% at 480 ms and 5.7% at 2,400 ms. That average uses character error rate for Japanese, Korean and Chinese and word error rate for the rest. It beats GPT Live Transcribe (8.4%) and Scribe v2 Realtime (8.7%). Gemini 3.5 Realtime Transcribe, at 4.2%, beats it on the same chart.

How did Voxtral 3 do on real voice orders?

In a smaller real-world test, both batch models recovered 74% of ordered products from real voice notes, ahead of Whisper large-v3 and behind Gemini 3.8 Flash. Reflekt Lab ran the three transcription models the day after launch on 35 WhatsApp voice notes, covering 18 orders placed with a grocery-delivery service in the UAE, spoken in Gulf Arabic, English and Urdu. It set no language and gave no prompt.

Source: Reflekt Lab, 22 October 2026.

The cheaper Mini model matched Small product for product and ran faster. Voxtral never got stuck repeating itself, which Whisper did on two notes. Two weaknesses showed up: Urdu came out in Devanagari (Hindi) script, and the realtime model sometimes wrote short notes in the wrong language. Reflekt Lab stresses that 31 products from one business in one city is a small sample.

How does Voxtral 3 speech generation work?

Voxtral TTS 3 reads the emotion of a passage from its text and adjusts tone, pace and emphasis without being told. Where you want control, you add inline tags such as [emotion=calm] or [pace=slow]. There are 34 emotions. The tags guide lists five pace settings, from very slow to very fast, but the limits page and the model card name only slow and fast, so test the others before you rely on them. Mistral says the tags work in English and calls them untested in French, German, Spanish, Italian, Portuguese and Hindi. In Voxtral 2, each preset voice came in fixed emotional variants; now one voice can move between moods in a single passage.

Mistral’s language guide lists 35 languages and more than 50 regional dialects, though the TTS model card still says nine. Mistral’s voice library lists 36 new preset voices covering 21 languages, including seven British English and four American English voices. The four Voxtral 2 speakers still work. For other languages, you clone a voice from a recording in that language. The model card quotes roughly 90 ms to first audio when streaming and does not say how it was measured. Mistral’s latency guide notes that the network and your player’s buffering add to what a listener hears, so measure the wait from your application. Output comes in five formats: MP3, WAV at 24 kHz, PCM, FLAC and Opus. Contextual TTS can also shape the delivery of a reply using audio from earlier in the conversation.

There are two tiers of cloning. Standard cloning works zero-shot from a reference clip, and you can save the result as a reusable voice ID. Saved voices carry a retention_notice setting, 30 days by default and adjustable from 1 to 365; the guide does not say whether a voice is deleted when it runs out. Professional cloning, which the announcement calls Pro Voice, is meant to keep a speaker’s character and emotional range across languages. Mistral’s guide recommends 30 minutes to three hours of clean audio from a single speaker, and publishes neither a price nor access terms.

The update also targets the failures that hurt customer service calls: numbers, dates, email addresses and unusual names. Mistral’s samples show Voxtral 3 reading “10/10” as a date where an earlier Voxtral speech model read it as a fraction. A pronunciation control accepts phoneme hints for rare or invented words. An expressiveness setting trades voice similarity for stronger delivery. Free-text style instructions are documented, but the instruction parameter is not supported on the current checkpoint.

Mistral’s speech benchmarks plot quality against price, and Voxtral sits at the cheap end with competitive scores. Most are internal evaluations scored by an AI judge (Gemini 3.8 Flash); one uses BeSimple AI’s Voice Code Bench, 300 English prompts built around numbers and names. Those are Mistral’s own task scores. In Artificial Analysis’s separate blind English preference arena, Voxtral 3 TTS ranks 12th of 98 entries, with an Elo score of 1,219, behind Eleven v4 (1,332) and Cartesia Sonic 3.6 (1,292). With cloned voices it ranks 13th, again below both. It does cost far less: $13 per million characters, against $49 for Sonic 3.6 and $80 for Eleven v4.

How does the Voxtral voice agent SDK work?

The SDK runs a whole conversation as one streaming pipeline. The realtime transcriber listens and spots when the caller has finished. A language model, such as Mistral Small 4, decides what to say and calls tools. Voxtral TTS speaks the answer in a tone that fits the conversation so far. Callers can talk over the agent, and it stops.

Mistral says each turn completes in under a second because the stages overlap, and gives no test method. Its latency guide measures a turn from the moment the caller stops speaking to the moment the agent starts answering, and describes the stages adding to that wait one after the other. Measure the delay on a real phone line before you rely on the claim.

The SDK is built on LiveKit, the open-source framework for live audio, though you do not need to know LiveKit to use it. It comes in Python and Node.js, and in TypeScript for the browser, where your server creates the session so the API key never reaches the page.

Agents can act during a call. Built-in web search, document libraries and workspace connectors run on Mistral’s side; your own functions run in your application. You can attach a short spoken filler to a tool to cover the pause while it runs; a tool without one stays silent. On a phone line none of your code is in the call, so client-side tools fail there, and anything a phone agent needs must be a server-side tool.

An agent is a saved, versioned configuration: instructions, models, voice, tools, welcome message and interruption rules. You can pin a version in production while testing the next. Signed webhooks tell your backend when a call starts, ends or fails, which is where you would store transcripts. The end-of-call transcript is best effort, though: if it is not ready in time, the webhook’s transcript field is null.

The agent works in browsers and apps over WebRTC and answers inbound phone calls over SIP. It cannot place calls or transfer a caller to a person yet, and SIP telephony is open only to enterprise customers through their Mistral account team. Each stage is a separate model, so you can swap the language model, voice or language without rebuilding the agent. Teams that want full control can skip the SDK and wire realtime transcription, a chat model and streamed speech together themselves. Our reviews of Grok Voice 3.0 and Gemini 3.8 Live and Gemini 3.8 TTS cover two of the hosted voice models Voxtral 3 competes with.

How much does Voxtral 3 cost?

Voxtral 3 transcription costs $0.0035 to $0.008 a minute, speech $13 per million characters and a hosted agent $0.04 or $0.08 a minute, at Mistral’s US dollar list prices. Against the previous generation, speech generation got 19% cheaper. Batch transcription on Mini rose by about 17% and realtime transcription by a third, while Small held at $0.004 a minute. The table sets Mistral’s list prices against the previous generation.

Sources: Voxtral 3 announcement, Mistral API pricing, Voxtral Transcribe 2 announcement, Voxtral TTS announcement, checked 24 October 2026. The pricing page bills the agent per minute of conversation without saying what that covers.

The choice of reasoning model doubles the agent price. At 10,000 call minutes a month, the agent API charge comes to $400 with Mistral Small 4 and $800 with a larger model. Mistral’s price list leaves three things unclear: whether the agent rate includes transcription and speech, what Pro Voice costs, and what on-premises deployment or phone lines cost. Ask before you budget a pilot.

Is Voxtral 3 open source?

No. Mistral’s Voxtral collection on Hugging Face lists no Voxtral 3 model, and the announcement links none. Self-hosting is possible, but only through Mistral’s sales team.

That is a change of course. Each earlier Voxtral release put at least one model on Hugging Face.

Mistral’s pitch is sovereignty instead: an EU endpoint, or the full stack on your own servers. A team that needs on-device or air-gapped speech recognition today can still use the open Voxtral Realtime model from February, or Voxtral Mini 4B Realtime Arabic, released under Apache 2.0. Whether Mistral will open any Voxtral 3 model later is unknown.

What changes when you migrate from Voxtral 2?

Voxtral 3 drops word-level timestamps and locks realtime settings on the first audio chunk, so existing code can break even if you only swap model IDs. Mistral’s migration guide maps each old model to its replacement.

The changes to check:

• Word timestamps are gone. Requesting word granularity now returns a 400 error; only segment timestamps remain. Subtitle and search tools that align single words will need rework.
• Realtime settings lock on the first audio chunk. Send every setting in the first session.update; a later update returns an error.
• Diarised segments can overlap in time. Displays that assume one speaker at a time need to show both, although Mistral says two people talking at once is not supported in this version.
• Emotion moves into tags. Voxtral 2 presets such as gb_jane_sad still work, but one voice can now carry several moods through [emotion=...] tags.
• Re-measure latency from your own application after the switch.

Two further traps sit outside the migration guide. Offline diarisation now requires segment timestamps, and a diarised request without them fails with a 422 error. And although the guide maps voxtral-small-2507 to Small Transcribe 3, the new model handles transcription only at launch: code that sends audio to chat completions for summaries, translation or questions should stay on voxtral-small-latest for now.

The voxtral-mini-tts-latest alias will point to Voxtral 3 within a month of general availability, so code that uses it will switch models without any change on your side. Pin a versioned ID if you need stable behaviour.

What should you check before building on Voxtral 3?

Test Voxtral 3 on your own calls, and confirm the phone and hosting terms with Mistral before a pilot. It is in public preview, and the parts a contact centre depends on carry the most limits: phone calls are inbound-only and enterprise-only, and on-premises pricing sits behind a sales conversation.

A sensible first test looks like this:

1. Pick one narrow, frequent call type, such as rescheduling an appointment or checking an order.
2. Load your product, place and staff names into context biasing, then transcribe real recordings with your callers’ regional accents.
3. Run the same recordings through your current provider and compare word error rates and diarisation yourself.
4. Time full turns on an actual phone line, not in the browser playground.
5. Listen for numbers, dates, email addresses and names in the generated speech, since that is where Mistral says it improved most.
6. Confirm retention settings and endpoint region with your privacy officer before any live calls.

Request rates and monthly quotas are set per workspace and shared by all its API keys, so check yours before any load test.

Read the hosting promises closely too. The global endpoint makes no commitment about where inference runs. The EU endpoint uses data centres in EU and EFTA countries and excludes stateful features such as Agents; account data, keys and billing may still be handled outside the region. Regional inference carries a 10% surcharge on token and cache charges, and the page does not say how that applies to per-minute or per-character audio prices. Zero data retention is a separate control for paid plans, which Mistral reviews and may approve or deny on request. It covers the speech and transcription endpoints but excludes Agents, and Mistral’s documentation does not say where the new Voice Agents fit on either list. Opting out of model training is a separate control again and does not need zero data retention. Our Mistral Large 4 review covers the same EU endpoint for Mistral’s text models.

Two legal points apply in Europe. For covered deployments, the EU AI Act’s transparency rules (Article 50) require that people be told they are talking to an AI system unless that is already obvious, so say it at the start of the call. Under Article 2, the Act can reach UK providers and deployers when the system’s output is used in the EU. Mistral’s own terms forbid cloning a voice without explicit consent; getting that consent in writing gives you a record of it.

Sources

Mistral AI

• Voxtral 3 announcement
• Voxtral announcement, July 2025
• Voxtral Transcribe 2 announcement
• Voxtral TTS announcement
• Changelog
• API pricing
• Audio overview
• Speech to Text overview
• Diarize speakers
• Text to Speech overview
• Emotion and pace tags
• Voices
• Voice cloning
• Style instructions
• Voice agents
• Telephony (SIP)
• Webhooks
• Languages
• Limits
• Latency
• Migrate from Voxtral 2
• Voxtral Small Transcribe 3 model card
• Voxtral TTS 3 model card
• Regional inference
• Zero data retention
• Additional Product Terms, section 11

Independent tests

• Speech to Text leaderboard (non-streaming), Artificial Analysis
• Provider Voice leaderboard, Artificial Analysis
• Controlled Voice leaderboard, Artificial Analysis
• Voxtral 3: Models, Pricing and a Real-World Test, Reflekt Lab

Coverage

• Mistral’s Voxtral 3 voice agents answer phone calls on EU-hosted servers, DataNorth
• Mistral launches Voxtral 3, Slator