Mistral Unveils Voxtral 3, Ranked 3rd Globally in Transcription

The new Voxtral 3 family brings together transcription models and a SDK designed for voice agents capable of reasoning and tooling. Artificial Analysis ranks Voxtral Small Transcribe 3 third globally for Speech to Text. The offering can be executed via an API hosted on European infrastructures or on clients' own servers, and it covers over 50 languages with speaker identification.

European API, on-premises execution, and announced customization

Mistral emphasizes the sovereignty of its solutions. Clients can access the models via the Mistral API, which relies on European infrastructures, or choose to run the models on their own servers. Companies have customization levers to tailor the entire system to business needs. The company describes agents supporting dozens of languages, equipped with rich vocal expression and complex reasoning capabilities, with deep integration into existing tools and data. According to Mistral, these agents interpret nuanced intentions, handle sophisticated requests, and can chain multi-step actions across connected systems. The language support includes around 50 languages, including 24 official languages of the European Union, as well as many low-resource Asian and Indian languages.

Demonstrated voice chain: transcription, reasoning, synthesis

The Voxtral 3 range includes Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, and a SDK that features a voice assistant presented as capable of performing complex reasoning and utilizing tools. During a demonstration, the artificial intelligence leverages the transcription provided by Voxtral to identify exchanges, calls upon the Mistral Large model for reasoning, and then produces the response using a vocal model. The video sequence illustrates that the assistant can be interrupted during its response and recognizes two distinct individuals by their voices. Mistral presents these components to build natural conversational experiences, contact center agents, and support services for virtual assistants and hands-free co-pilots.

Multilingual transcription and third global rank according to Artificial Analysis

Mistral highlights the transcription performance of Voxtral 3. Voxtral Small Transcribe 3 supports over 50 languages and can distinguish between speakers, a feature presented as suitable for meeting minutes. For speech-to-text conversion, Artificial Analysis ranks this model as the third-best AI in the world, noting a very low error rate. In this evaluation framework, Mistral's technology outperforms those of Google, SpaceX, and OpenAI.