Voxtral 3
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Mistral has announced Voxtral 3, a family of speech-to-text and text-to-speech models paired with a voice-agent SDK for real-time conversations and tool use. The company says the models support transcription in 50 languages and speech generation in more than 35; its performance comparisons are based on benchmarks published by Mistral. The products are available through Mistral’s EU-hosted API or for self-deployment, but independent verification and full deployment details are not provided in the supplied material.

Mistral has announced Voxtral 3, a family of speech-to-text and text-to-speech models and a voice-agent SDK designed to handle real-time conversations and actions across connected services. The release matters to organizations building voice assistants and contact-center systems because Mistral says it can combine transcription, generated speech and tool use in one streaming pipeline, with access through an EU-hosted API or deployment on customers’ own infrastructure.

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

The release includes Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 for speech recognition, as well as Voxtral Mini TTS 3 for speech generation. Mistral says transcription support covers 50 languages, including 24 official EU languages, and includes speaker diarization, end-of-turn detection and context biasing. Context biasing lets an application provide terms or other text that may help the model recognize specialized vocabulary; Mistral says this can be done without fine-tuning.

Mistral describes the SDK as a way to connect listening and speech generation with reasoning and tool calling, so an agent can retrieve information or interact with business systems during a conversation. The company says the text-to-speech model supports more than 35 languages and 50 dialects, with voices that adapt emotion, pace and emphasis to conversational context. A separate Pro Voice feature is described as adapting a speaker’s character and style across languages from 30 minutes of audio.

Mistral says Voxtral 3 is available through its EU-hosted API or for self-deployment. It says those options let organizations control where audio is processed and whether conversations are retained. The supplied announcement also lists a price of $0.0035 per minute for the 4-billion-parameter Voxtral Mini Transcribe 3; it does not specify in the provided material whether that rate applies to every deployment or all usage conditions.

At a glance
announcementWhen: Announced by Mistral; the company says…
The developmentMistral announced Voxtral 3, combining multilingual speech models with a real-time SDK intended to let voice agents listen, reason, call tools and respond.

Voice Agents Move Beyond Transcription

Voice systems often rely on separate services for recognizing speech, deciding what to do and producing a spoken reply. Mistral is pitching Voxtral 3 as a way to bring those pieces together with live tool access, potentially allowing an agent to complete tasks rather than only answer questions. For businesses, that could make voice interfaces more useful in service desks, contact centers and hands-free applications, if the system performs reliably in real operating conditions.

The announcement also speaks to a practical concern for organizations handling sensitive audio: where recordings are processed and whether they are retained. API access hosted in the EU and self-deployment are offered as control options, though the announcement alone does not establish how a particular customer’s data will be handled. Buyers will need to compare those controls, performance and costs against their own legal and technical requirements.

Mistral’s headline benchmark results could attract developers evaluating multilingual transcription, but they are company-reported comparisons. Scores on curated tests do not by themselves establish performance on every accent, noisy call, specialist vocabulary or real-time workflow. Operational testing remains relevant before organizations rely on the models for customer interactions or consequential decisions.

Products Worth Considering

Sonos One Two-Room Set Smart Speakers with Built-In Alexa, Black

Sonos One Two-Room Set Smart Speakers with Built-In Alexa, Black

Jackery Explorer 1000 v2 Portable Power Station, 1024Wh LiFePO4 Battery

Jackery Explorer 1000 v2 Portable Power Station, 1024Wh LiFePO4 Battery

As an affiliate, we earn on qualifying purchases.

Models, Benchmarks and Availability

Mistral’s announcement frames Voxtral 3 as a response to the difficulty of making speech systems work as a coordinated agent: they must recognize words, interpret intent, respond naturally and carry out actions. Earlier speech tools commonly focused on transcription or generated speech alone. Voxtral 3’s distinguishing product claim is that its SDK links those stages in a real-time streaming system, alongside models for transcription and speech synthesis.

The company reports that Voxtral Small Transcribe 3 achieved a 4.0% macro-average word error rate in both forced-language and inferred-language settings on the FLEURS 45-language benchmark. In the same table, Mistral lists 4.1% for Microsoft AI Transcribe 2, 4.5% for Gemini 3.5 Transcribe in the inferred setting and 14.9% for Whisper large-v3 in that setting. The comparison covers the 45 languages supported by all seven systems in the table; language-specific scoring exceptions use character error rate rather than word error rate.

Mistral also reports an 18.4% diarization cpWER for Small Transcribe 3 across five public benchmarks, compared with 18.6% for Voxtral Mini Transcribe 3 and 20.3% for Microsoft AI Transcribe 2. These are results reported by the vendor, not an independent evaluation in the source material. Mistral says the Mini transcription model has 4 billion parameters and is priced at $0.0035 per minute. The announcement does not provide equivalent pricing for every model or deployment option.

“A reliable voice agent needs to do four things very well (and simultaneously): Listen accurately across languages, accents, speakers, and interruptions, Reason deeply about what users mean in specific contexts, Speak naturally with emotion and expression, and Take action over multiple steps through connected systems.”

— Mistral, in its Voxtral 3 announcement

Independent Tests and Deployment Details

The supplied announcement does not include independent evaluations of Voxtral 3’s transcription, voice quality or complete agent workflow. Its tables report vendor-selected benchmark comparisons, and results in those tests may not predict performance for every language, accent, recording condition or business task. The announcement does not provide enough detail to establish how the integrated SDK performs under production traffic or what latency users should expect.

Some practical terms also remain unspecified in the material: pricing beyond the stated Mini Transcribe 3 rate, model-by-model availability, service limits, the exact retention defaults for API use and the technical requirements for self-deployment. Mistral describes voice adaptation using 30 minutes of audio, but the source does not explain consent controls or other safeguards around reproducing an individual’s voice. Customers will need product documentation and contractual terms to answer those questions.

Customer Testing Will Set the Benchmark

Mistral says Voxtral 3 is available now through its EU-hosted API and for self-deployment. The next test is whether developers and organizations can reproduce the reported results with their own audio, languages and workflows, including interruptions, specialist terms and calls that require multiple tool actions.

Potential adopters should also check the full pricing and data-handling terms for their chosen deployment, and evaluate the speech generation and voice-adaptation features with appropriate consent and oversight. The announcement does not identify a later launch date or a specific upcoming release milestone. Further information is likely to come through Mistral’s product documentation and customer testing, but the timing and details of any updates have not been stated.

Key Questions

What is Voxtral 3?

Voxtral 3 is Mistral’s family of speech-to-text and text-to-speech models, released alongside an SDK designed to build real-time voice agents that can use connected tools.

Which languages does it support?

Mistral says the transcription models support 50 languages. It says Voxtral Mini TTS 3 supports more than 35 languages and 50 dialects. The supported language sets differ by feature.

How much does Voxtral 3 cost?

The announcement lists $0.0035 per minute for Voxtral Mini Transcribe 3. It does not give full pricing for all models or deployment options in the supplied material.

Can organizations run Voxtral 3 on their own infrastructure?

Yes. Mistral says customers can use its EU-hosted API or self-deploy the models on their own infrastructure. Specific technical requirements and data-retention terms are not included in the source material.

Are the benchmark claims independently verified?

The supplied results are reported by Mistral. The source material does not include an independent evaluation, so customers should treat the figures as vendor-published benchmark results and test their own use cases.

Source: hn

You May Also Like

Razer Surges In Global Coverage

Razer is experiencing a significant increase in international media coverage, with 14 mentions in recent reports, indicating rising global interest.

Technology Operations Signal Monitor: Cool URIs Don’t Change (1998)

An IdeaNavigator AI proposal would test whether small-company product and engineering leads value role-filtered briefs on technology changes.

How To Choose From The 10 Best Smartwatches For 2027

A 2027 smartwatch buying guide compares Apple, Garmin and budget models, with battery life, phone compatibility and health tracking as key tradeoffs.

Zenless Zone Zero Enters The Steam Most-played Chart

The game has reached the Steam most-played chart, peaking at over 10,000 players, marking a significant rise in its player base.