Mistral AI launches Voxtral 3 voice agent stack with real-time tool use

Mistral AI has launched Voxtral 3, a new family of language models specifically engineered for real-time voice agents. According to a report by The Decoder, the release marks the introduction of a comprehensive stack that integrates listening, thinking, and speaking capabilities into a unified system for developers and enterprises.

The Voxtral 3 suite includes three primary components: Voxtral Small Transcribe 3 for speech-to-text conversion, Voxtral Mini TTS 3 for text-to-speech synthesis, and a voice agent SDK based on the LiveKit framework. These elements work together to create a single streaming pipeline that handles audio input, processes intent, and generates audio output without significant latency. This architecture allows agents to perform complex tasks while maintaining continuous voice interaction with users.

Technical capabilities and language support

The transcription component of Voxtral 3 covers more than 50 languages, enabling broad global accessibility for voice-driven applications. The text-to-speech module supports over 35 languages and approximately 50 distinct dialects, with features designed for emotional adaptation in the generated voice output. This level of linguistic coverage allows organizations to deploy voice agents in diverse markets without requiring separate model configurations for each region.

Beyond basic translation and synthesis, the system enables agents to interact with external digital tools during live conversations. Mistral AI states that agents can tap into external systems to perform actions such as booking appointments or checking invoice statuses. This tool-use capability transforms the voice agent from a passive responder into an active participant that can execute transactional tasks while speaking with a user.

Deployment options and data control

Mistral AI offers flexible deployment models for Voxtral 3 to address data sovereignty and privacy concerns. Organizations can run the system on their own infrastructure, ensuring that sensitive audio and conversation data never leaves their private networks. Alternatively, companies can utilize EU-hosted APIs, which keep data processing within European jurisdictional boundaries. These options are designed to help enterprises maintain control over their data while leveraging advanced voice AI capabilities.

Pricing for the transcription service starts at $0.0035 per minute, according to The Decoder. Detailed technical specifications and integration guidelines are available in the official documentation provided by Mistral AI. The launch positions Voxtral 3 as a competitive option for developers seeking to build sophisticated, real-time voice interfaces with robust language support and tool integration.

What should you do?

Security and technology professionals evaluating voice AI integrations should assess the data handling practices of new deployments. With the ability to access external tools, voice agents represent a potential vector for unauthorized actions if not properly secured. Teams should review how the agent authenticates with external systems and what permissions are granted during live conversations.

What to do and how to stay safe: Mistral AI

• Audit the permissions granted to voice agents when they access external tools like booking systems or financial databases to ensure least-privilege access.
• Verify that deployment configurations, whether on-premises or via EU-hosted APIs, align with your organization's data residency and privacy compliance requirements.
• Monitor API usage and conversation logs for anomalous patterns that could indicate unauthorized tool usage or prompt injection attempts.
• Review the LiveKit-based SDK documentation for security best practices related to real-time streaming connections and session management.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What components are included in the Voxtral 3 stack?

Voxtral 3 includes Voxtral Small Transcribe 3 for speech-to-text, Voxtral Mini TTS 3 for text-to-speech, and a voice agent SDK based on LiveKit.

How many languages does Voxtral 3 support?

The system supports transcription in more than 50 languages and text-to-speech in over 35 languages with 50 dialects.

Can Voxtral 3 agents interact with external systems?

Yes, agents can tap into external tools during live conversations to perform tasks such as booking appointments or checking invoices.