Mistral Voxtral 3 Review: Voice Agent SDK & Pricing

Mistral describes Voxtral 3 as a new family of state-of-the-art speech-to-text and text-to-speech models, plus a voice agent SDK that brings listening, thinking, tool calling and speaking into a single real-time streaming pipeline. The target use cases are contact-centre agents, service desks, virtual assistants and hands-free copilots.

• Transcription: Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3, covering 50+ languages, with diarisation (who said what), context biasing and end-of-turn detection.
• Realtime: Voxtral Mini Realtime 3, built for agentic use at sub-500ms latency in over 35 languages.
• Speech: Voxtral Mini TTS 3, in batch and realtime, with emotion, pace and emphasis control across 35+ languages and 50+ dialects.
• SDK: a voice agent SDK combining listening, reasoning, tool calling and speaking over phone or web.
• Deployment: Mistral's EU-hosted API, or self-deployment on your own infrastructure.

Our verdict in one line: this is the most complete European answer yet to the voice-agent stack, with credible published numbers and aggressive pricing. It is also a launch, which means the headline claims are Mistral's, the model weights situation is murky, and real-world call-centre performance remains to be proven.

The family has several distinct parts, and the naming is easy to muddle. Mistral's announcement lists Voxtral Small Transcribe 3 as its flagship speech-to-text model. Voxtral Mini Transcribe 3 is the 4-billion-parameter variant. Voxtral Mini Realtime 3 handles streaming. Voxtral Mini TTS 3 does speech generation, and the voice agent SDK ties the parts together.

Mistral describes the speech-to-text model as state-of-the-art in multilingual transcription and as capable of diarisation, using keyphrases to improve accuracy (context biasing) and detecting when someone has finished speaking so an agent can respond without delay (end-of-turn detection). Those last three features are the difference between a transcriber and something you can build a conversation on.

The announcement page is dated 22 October 2026 in its metadata, and secondary reports place the synthesis model's public preview on the Mistral API on 20 October 2026. It follows the original open-weight Voxtral line of July 2025, and there is a prior Mistral TTS release earlier in 2026, so this is the third-generation framing of an established product line.

Mistral says the Small model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe, and the Mini variant is a 4B-parameter model priced at $0.0035 per minute. Both support context biasing, where key terms and free-form text are supplied to improve recall on technical vocabulary without fine-tuning. The supported list expands to 50 languages, including 24 official EU languages and many low-resource Asian and Indic languages.

Word error rate

On the FLEURS benchmark (45 languages supported by all seven models compared), Mistral charts macro-average word error rate in two settings: forced, where the language is specified, and inferred, where it is auto-detected. The Voxtral models sit at the low end of the chart, with the lowest plotted figures around 4.0% to 4.1% for the leading models, and Whisper large-v3 at the high end around 16.0% and 14.9%. Mistral also cites Artificial Analysis's word error rate index, a weighted average over VoxPopuli, Earnings-22 and AgentTalk, where Voxtral Small Transcribe 3 ranks near the top of a long list of commercial and open models.

Treat the exact ordering with care. Mistral selected the comparison set and the charts, and the margins between the top few models are small enough that a different language mix could reorder them. The consistent result, Voxtral at or near the front and Whisper large-v3 well behind, is more reliable than any single decimal place.

Diarisation

Diarisation is where voice agents quietly fail: if the system cannot tell the customer from the agent, summaries and actions go wrong. Mistral reports concatenated minimum-permutation word error rate across public benchmarks including AMI, Switchboard, CallHome and AISHELL-4. The Voxtral models post the lowest figures on its chart, ahead of Pyannote Orchestrator, Microsoft AI Transcribe 2, Scribe v2 and Gemini 3.5 Transcribe, with the weakest compared model above 38%. Mistral notes it used 30-minute versions of AMI and AISHELL for compatibility.

Context biasing

Context biasing is the feature enterprises will notice. On Earnings-22 keyphrase recall, Mistral's chart shows Voxtral Small Transcribe 3 with context recalling roughly 94.6% of keyphrases against 81.4% without, ahead of Scribe v2 and Gemini 3.5 Transcribe in the same test. In plain terms, telling the model your product names and speaker names materially reduces the chance it mangles them. If your calls are full of drug names, part numbers or surnames, this matters more than a half-point of general word error rate.

Voxtral Mini Realtime 3 is designed for agentic use at sub-500ms latency in over 35 languages, with diarisation and context biasing. Mistral's FLEURS 35 realtime chart compares it at several latency settings (240, 480, 960 and 2400 ms) against realtime systems including Gemini 3.5 Realtime Transcribe, Microsoft AI Transcribe 2 Streaming, GPT Live Transcribe and Scribe v2 Realtime. The offline Voxtral Mini Transcribe 3 is shown as an offline baseline, and the realtime models trade a little accuracy for latency, as you would expect.

End-of-turn prediction is natively supported, using both acoustic and semantic information. That is the practical heart of conversational feel: cut a customer off and they get angry; wait too long and the call feels dead. Our earlier coverage of OpenAI's GPT-Live turnless voice system looked at the same problem from the other direction, with a model that removes turn-taking altogether. Voxtral takes the more conventional pipeline route but invests heavily in the turn-detection part.

Secondary reports put the realtime model's batch price at about $0.008 per minute, but we could not find that figure on the official page, so treat it as unverified.

The text-to-speech model produces voices that adjust emotion, pace and emphasis according to conversational context, with 35+ languages and 50+ dialects. Mistral says it interprets the semantic meaning of context natively, so a sentence that should sound apologetic sounds apologetic without explicit instructions. For explicit control it exposes emotion, pacing and emphasis as variables, supports non-standard phoneme hints for unusual names, and offers contextual TTS and a vocabulary library.

Voice cloning is split into two tiers. Standard cloning uses 30+ minutes of sampled audio, and a Pro Voice adaptation service aims to reproduce a speaker's distinctive character, emotional range and speaking style across languages, again from about 30 minutes of audio. Mistral positions Pro Voice for professional use cases where basic cloning does not suffice.

Price and quality

Mistral lists synthesis at $13 per million characters and says that is cheaper than comparable offerings from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. It backs the claim with two quality-per-dollar charts: Voice Code Bench, built by BeSimple AI around support failure cases involving numbers and unique names (300 English prompts, two voices per system), and an internal pronunciation-control test with 300 sentences each containing one rare or invented word. Both charts were generated on 20 October 2026, use a Gemini 3.8 Flash model as judge, and take prices from vendor pricing pages.

Two cautions. One chart is an internal evaluation, and an LLM judge is a proxy for human listening. Mistral also says systems without documented pronunciation control are excluded from the plot. If voice quality is the whole point of your product, run a blind listening test against your current provider, including ElevenLabs, on your own scripts before moving anything.

Mistral frames a reliable voice agent as four things that must work simultaneously: listening accurately across languages, accents, speakers and interruptions; reasoning about what users mean in context; speaking naturally with emotion; and taking multi-step action through connected systems. The SDK's job is to put those four in one streaming pipeline so that agents can retrieve information, interact with business systems and complete tasks during live conversations over phone or web.

One secondary report says the SDK is built on LiveKit, the open-source realtime media stack, though we could not confirm that on the official page and flag it as reported rather than established. If accurate, it would make integration with existing telephony and WebRTC deployments easier.

This puts Mistral in direct competition with the hosted voice-agent platforms and with consumer products such as Hermes Apollo voice mode. The difference is the sovereignty story: the pipeline can run in Mistral's EU-hosted environment or on infrastructure you control, so audio does not have to leave your jurisdiction.

Mistral stresses that, like its other models, Voxtral 3 is designed for sovereign deployment: available through its EU-hosted API or for self-deployment, with organisations controlling where audio is processed, how data is handled and whether conversations are retained. For UK and EU businesses handling regulated calls, such as health, financial services and public-sector contact centres, that is a more important feature than a decimal of word error rate.

The open-weights picture is unclear. Mistral says the family can be self-deployed, but one report describes Voxtral 3 as closed-weight, a departure from the Apache 2.0 licence of the original July 2025 Voxtral, and another says the synthesis model's weights are not published. Both can be true at once: self-deployment may mean a commercial licence for weights you receive rather than a public download. Read the licence before you plan an air-gapped deployment.

For anyone who valued the original open-weight Voxtral, that is a real change to weigh. For anyone who simply wants European data residency and predictable pricing, it may not matter.

Mistral lists prices in US dollars. We have not converted them to pounds because exchange rates move, so check the current rate before budgeting.

• Voxtral Mini Transcribe 3: $0.0035 per minute, which works out at $0.21 per hour of audio.
• Voxtral Mini TTS 3: $13 per million characters. At roughly 15 characters per second of speech, a typical figure for English, a million characters is about 18 hours of audio, though the rate varies by language and voice.
• Realtime transcription: secondary reporting says about $0.008 per minute; unverified on the official page.
• Small Transcribe 3: a price was not listed in the material we reviewed.

A worked example: a contact centre handling 10,000 five-minute calls a month generates 50,000 minutes of audio. At $0.0035 per minute, transcription with the Mini model costs about $175 a month. Even if you doubled that for real-time and quality headroom, the transcription line item would be small next to staffing, which is exactly why this pricing will attract attention.

On Mistral's own charts, Voxtral Small Transcribe 3 leads or ties for the lead against Whisper large-v3, Scribe v2, Gemini 3.5 Transcribe, Microsoft AI Transcribe 2 and GPT Transcribe on multilingual and diarisation tests. Artificial Analysis reportedly ranked it third on its speech-to-text leaderboard, which suggests the claims survive at least one independent comparison, though not a clean sweep.

On TTS, Mistral claims a cost advantage against OpenAI, Deepgram, ElevenLabs Flash and Gemini Flash, and a quality-per-dollar advantage on its two tests. Whether a listener prefers the voices is a subjective question no chart can settle.

Strategically, the release sits alongside the push into agentic assistants we have covered elsewhere, such as Claude Duet, which acts on its own, and Mistral's own large-model line in Mistral Large 4 (Le Chonk). Voxtral is the audio layer that any of those assistants would need if you want them to take phone calls.

Benchmarks tell you where to look; your own audio tells you what to buy. The features Mistral has emphasised map neatly onto a handful of workloads, and each suggests a cheap first experiment.

Contact centres and service desks

The obvious fit is the call that follows a script but is full of names, numbers and accents. Diarisation separates agent from customer, context biasing protects product names, and end-of-turn detection stops the system talking over people. A sensible pilot is to replay a few hundred anonymised recordings through the Small model with and without a vocabulary list, then compare errors on the terms that matter to your business rather than on overall word error rate.

Meetings, interviews and research audio

Anyone transcribing multi-speaker audio will care most about who said what. Run a handful of recordings that include crosstalk and a poor microphone, and check speaker attribution by hand. The Mini model's price makes it cheap enough to transcribe an entire archive as a trial, though you should still sample for errors before relying on summaries built on top.

Voice assistants and hands-free copilots

Here latency and tone dominate. Measure end-to-end response time, not just transcription speed, and listen for the awkward pauses and cut-offs that destroy trust. Compare the expressive TTS blind against your current voice, and include the hard cases: product names, numbers read aloud and sentences that should sound sympathetic.

A short checklist before you commit

• Run at least 100 real, anonymised calls or recordings through each candidate, in the accents and noise conditions your customers actually produce.
• Score errors on business-critical terms separately from overall accuracy.
• Time the full round trip from the end of a customer's sentence to the start of the reply.
• Confirm in writing where audio is processed, how long it is retained and what the licence allows if you self-deploy.
• Ask what consent, watermarking and abuse controls apply to voice cloning before enabling it.
• Model your monthly minutes and characters at the listed prices, then add a margin for retries and peak load.

• Vendor-selected benchmarks. The comparison sets, languages and judges were chosen by Mistral or its partners. Independent leaderboards will matter.
• LLM-judged speech quality. Both TTS charts use a language model as judge. It is a useful proxy, not a replacement for human listening tests.
• Language claims conflict. Mistral says transcription covers 50+ languages while TTS covers 35+ languages and 50+ dialects; some reports blur the two.
• Weights and licence. Conflicting reports about open versus closed weights need resolving before any self-hosting plan.
• Voice cloning risk. Cloning from 30 minutes of audio is useful for brands and equally attractive to fraudsters. Ask what consent and watermarking controls ship with the product.
• No real-world call data. Published results are benchmarks. Noisy phone lines, crosstalk and accents are where voice agents fail.

Try it now if you run a multilingual contact centre, need European data residency, transcribe domain-heavy audio such as earnings calls or clinical notes, or want to cut TTS costs without moving to a self-built stack. Context biasing and diarisation make it a sensible first test for meeting and call analytics.

Wait if you depend on open weights, need a published safety and misuse policy for voice cloning before launch, or have a voice product where subjective voice quality is the differentiator and you have not yet run a blind comparison.

Voxtral 3 is a serious, well-priced, full-stack entry in voice AI that leans on Europe's data-sovereignty advantage. The published numbers put it at or near the front of multilingual transcription and diarisation, and the TTS pricing undercuts the established names. The caveats are the usual launch-week ones: vendor-chosen benchmarks, LLM judges, an unclear weights policy and no field data yet.

Run your own audio through it, compare it blind against what you use today, and read the licence before you commit. If the results hold, the cost of a voice agent just dropped again.