Mistral launches Voxtral 3 voice agent models and SDK

Article

Mistral launches Voxtral 3: connecting speech recognition, reasoning, and tool calling into a real-time agent

Beating AI news flash: French AI company Mistral launches Voxtral 3, including speech transcription models, speech synthesis model, and a complete Voice Agent SDK.

In the official demonstration, a user asks in English to book an appointment with a dermatologist.AI reads a French referral letter, calls the system to check available slots, and confirms the appointment in English after booking it. The entire process is completed through a real-time voice conversation.

Voxtral 3 is not an end-to-end large voice model that handles every task. It uses a modular design: transcription models recognize speech and determine when the user has finished speaking; general-purpose large models such as Mistral Small handle reasoning and tool calling; and the speech synthesis model generates responses. The LiveKit-based SDK connects these stages into a real-time pipeline, supporting telephone and web calls and allowing users to interrupt at any time. Mistral claims that single-turn interaction latency is under 1 seconds. Developers can also independently replace the reasoning model or voice model.

The number of languages supported for transcription has expanded from the previous generation's 13 languages to more than 50, while speech synthesis supports more than 35 languages and more than 50 dialects. In Artificial Analysis's independent non-real-time transcription test, Voxtral Small Transcribe 3 word error rate was 2.1%, ranking third, behind only Alibaba's and Microsoft's models. Speech synthesis ranked 13。

API real-time transcription is priced per minute at 0.008 USD, and speech synthesis per million characters at 13 USD. Enterprises can also contact Mistral to deploy on their own servers or in a private cloud.

Original link https://m.theblockbeats.info/flash/372884

Key points

Voxtral 3 including speech transcription models, speech synthesis model, and a complete Voice Agent SDK。
In the official demonstration,AI reads a French referral letter, checks available slots, and books an appointment with a dermatologist, then confirms it in English.
The system uses a modular design, with transcription models, general-purpose large models such as Mistral Small, and speech synthesis models each handling different stages.
The LiveKit-based SDK supports telephone and web calls, allows users to interrupt at any time, and also allows developers to replace the reasoning model or voice model.
Mistral claims that single-turn interaction latency is under 1 seconds.
Transcription supports 50 languages, while speech synthesis supports more than 35 languages and more than 50 dialects.
Artificial Analysis's independent non-real-time transcription test shows that Voxtral Small Transcribe 3 word error rate was 2.1%, ranking third; speech synthesis ranked 13。
API Real-time transcription is priced per minute at 0.008 USD, and speech synthesis per million characters at 13 USD。
Enterprises can contact Mistral to deploy on their own servers or in a private cloud.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is not a voice model that independently completes all tasks, but rather a pipeline with distinct roles: it first converts speech into text, then a general-purpose model reasons and calls system tools, and finally synthesizes the response into speech.Voice Agent SDK is responsible for connecting these stages, enabling developers to build interruptible real-time telephone or web voice assistants.

Why it matters to readers

The modular design allows developers to replace reasoning models and voice models separately, and to integrate voice conversations into tasks such as appointment booking that require calls to external systems. The disclosed language coverage, test metrics,API pricing, and private deployment options also provide information for developers evaluating the solution.

Beginners can understand it as a toolkit for building real-time voice assistants, rather than a single model that handles everything. Key considerations include recognition accuracy, response speed, tool-call reliability, and cost of use.

Risks and unknowns

Less than 1 seconds of single-turn latency is Mistral's claim; the evidence does not provide independent real-time retest results.
2.1% word error rate comes from independent non-real-time transcription tests and cannot directly represent real-time voice conversation performance.
The evidence only states that speech synthesis ranks 13, without providing the corresponding score or test details.

Mistral launches Voxtral 3 voice agent models and SDK

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center