Voxtral 3

TL;DR

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.Hide

Mistral has announced Voxtral 3, a family of speech models and a voice-agent SDK designed to combine transcription, reasoning, tool use and speech in real time. The company says the models cover more than 50 transcription languages and can be accessed through its EU-hosted API or self-deployed; performance comparisons in the announcement are based on Mistral-reported benchmarks.

Mistral has announced Voxtral 3, a family of speech-to-text and text-to-speech models paired with a voice-agent software development kit intended to handle live conversations, tool calls and multi-step tasks. The company says Voxtral 3 is available through its EU-hosted API or for deployment on customer infrastructure, giving organizations options for where audio is processed.

The release includes Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 for transcription, alongside Voxtral Mini TTS 3 for speech generation. Mistral says transcription support spans 50 languages, including 24 official EU languages, and includes speaker diarization, contextual prompting for domain-specific terms, and end-of-turn detection. The latter is designed to help an agent recognize when a speaker has finished so it can respond.

Mistral says Voxtral Small Transcribe 3 performs strongly on multilingual transcription and diarization benchmarks. In the company’s reported FLEURS comparison, its macro-average word error rate across 45 languages was 4.1% when the language was specified and 4.0% when it was inferred. The benchmark covers seven models and uses character error rate rather than word error rate for Japanese, Korean, Thai and Chinese. Lower scores indicate fewer recognition errors.

The company says Voxtral Mini Transcribe 3 has four billion parameters and costs $0.0035 per minute. For speech generation, Voxtral Mini TTS 3 supports more than 35 languages and 50 dialects, with voice delivery that can vary emotion, pace and emphasis. Mistral also describes Pro Voice adaptation, which it says can reproduce a speaker’s voice characteristics and speaking style across languages using 30 minutes of audio.

Voice Agents Move Beyond Transcription

The release targets a persistent gap between speech recognition and a useful conversational agent. A system that transcribes accurately still needs to interpret what a person wants, retrieve relevant information, use business software and respond naturally. Mistral’s SDK is presented as one pipeline for those steps, which could make voice agents more practical for contact centers, service desks and hands-free assistants.

Deployment choices may matter to organizations handling sensitive conversations. Mistral says customers can use an EU-hosted API or run the models on their own infrastructure, and can control whether audio is retained. Those options speak to data location and operational control, but the announcement alone does not establish how the product performs under every organization’s privacy, security or compliance requirements.

The pricing claim for Mini Transcribe 3 gives buyers a starting point for estimating transcription costs, but it does not describe the full cost of an operating voice agent. Speech generation, reasoning, tool integrations, infrastructure and service volume can add expenses. Practical value will depend on accuracy in a given setting and whether the agent can complete tasks reliably rather than only produce fluent responses.

How Voxtral 3 Fits Together

Mistral frames Voxtral 3 as a coordinated set of speech models and agent software, rather than a standalone transcription release. Its stated design joins listening, reasoning, calling connected tools and speaking in a real-time streaming experience. The intended applications include phone and web interactions, virtual assistants and copilots.

The transcription models can use context biasing: users provide key terms or free-form context to improve recognition of names and specialist vocabulary, according to Mistral. The company says this can help production use without fine-tuning. That feature is relevant where ordinary language models may mishear technical terminology, although the announcement does not provide results for every industry or acoustic condition.

Mistral’s benchmark tables compare Voxtral models with systems including Whisper, Gemini, Microsoft and Scribe. The published figures are company-reported test results, not an independent evaluation in this announcement. They provide a stated comparison under named benchmarks, but should not be read as a guarantee for performance on every language, accent, recording setup or live conversation.

“A new family of state-of-the-art speech-to-text and text-to-speech models, alongside a voice agent SDK that brings listening, thinking, tool calling, and speaking into a single real-time streaming pipeline.”

— Mistral, in its Voxtral 3 announcement

Independent Tests and Deployment Details

The announcement does not include independent verification of the benchmark comparisons or detailed testing across real customer deployments. Results in the FLEURS and diarization tables reflect specified datasets and scoring methods; performance in noisy settings, with overlapping speakers or across less-tested accents may differ. The release does not establish how the systems compare on end-to-end task completion, latency or error recovery in a live agent workflow.

It is also not clear from the supplied announcement which features, service limits or safeguards are available across every deployment option. Mistral says customers can control audio retention and self-deploy, but does not provide detailed retention defaults, security documentation, service-level commitments or the full pricing structure for the SDK and associated models here. The terms and consent requirements for voice adaptation using a person’s audio are not detailed in the announcement.

Customer Testing Will Show Practical Reach

Mistral says Voxtral 3 is available now through its EU-hosted API and for self-deployment. The next evidence to watch is how customers and independent evaluators test transcription quality, speech latency and agent task completion in actual contact-center and assistant workflows. Results across languages, accents and noisy calls will help establish whether the benchmark performance carries over to day-to-day use.

Organizations considering adoption will also need to review deployment documentation, audio-retention controls, voice-adaptation terms and total operating costs. Mistral has not outlined a later launch date or a separate product milestone in the announcement, so further updates on those details remain subject to the company’s releases.

Key Questions

What is Voxtral 3?

Voxtral 3 is Mistral’s family of speech-to-text and text-to-speech models, released with a voice-agent SDK intended to support live conversation, tool use and multi-step tasks.

How many languages does Voxtral 3 support?

Mistral says its transcription models support 50 languages. Voxtral Mini TTS 3 is listed as supporting more than 35 languages and 50 dialects.

Where can organizations run Voxtral 3?

Mistral says it is available through the company’s EU-hosted API or for self-deployment on customer infrastructure. The announcement does not provide every operational or service detail for those options.

Are Voxtral 3’s benchmark results independently verified?

The supplied announcement reports Mistral’s own benchmark comparisons. It does not include independent verification, so the figures should be understood within the datasets and scoring methods the company describes.

Source: hn