Mistral AI Launches Voxtral 3 Voice Agent Stack with Real-Time Tool Use
The new suite includes transcription, speech synthesis and an SDK for building low-latency voice agents in over 50 languages.
Key points
- Voxtral 3 features three transcription models, one TTS model and a LiveKit-based SDK.
- Small Transcribe 3 achieves a 2.1% AA-WER, ranking third on Artificial Analysis.
- The system supports real-time streaming with sub-second turn completion times.
Mistral AI has released Voxtral 3, a comprehensive voice agent stack that combines speech-to-text, text-to-speech and real-time interaction capabilities, according to MarkTechPost. The launch includes a family of transcription models, a synthesis model and a software development kit designed for building interactive voice applications across more than 50 languages.
The Voxtral 3 suite comprises three distinct transcription models, one text-to-speech model and a voice agent SDK built on LiveKit technology, per MarkTechPost. All four components are currently available in public preview via the Mistral API, allowing developers to access the tools through endpoints in the EU, US and globally.
View the original post on X
View the original post on X
Technical Specifications and Performance
Voxtral Small Transcribe 3 is the flagship batch transcription model, achieving a 2.1% AA-WER score and ranking third on Artificial Analysis, according to MarkTechPost. Voxtral Mini Transcribe 3 is a 4-billion-parameter model designed for resource-constrained environments, offering faster and cheaper batch processing while maintaining strong performance metrics.
The real-time streaming model, Voxtral Mini Transcribe Realtime 3, handles speech recognition over WebSocket connections with end-of-turn detection. In benchmarks, this model recorded a 7.4% WER on FLEURS 35 at 240 milliseconds, which MarkTechPost notes is higher than Gemini 3.5 Realtime’s 4.2% WER on the same test set.
All transcription models support speaker diarization and context biasing, allowing users to input key terms to improve recall rates. On the Earnings 22 dataset, the Small model achieved 94.6% keyphrase recall with context biasing enabled, compared to 81.4% without it, according to the report.
Agent Architecture and Use Cases
The accompanying SDK creates a unified streaming pipeline that integrates listening, reasoning and speaking into a single process. Realtime Transcribe 3 detects when a user stops speaking, after which a general large language model, such as Mistral Small 4, processes the input and triggers tool calls.
Voxtral Speech 3, also referred to as Voxtral Mini TTS 3, generates audio responses that match the conversation’s tone. Each conversational turn completes in under one second, and the agent automatically pauses if the caller interrupts, enabling natural dialogue flow.
The system supports phone calls via SIP and browser or app sessions through WebRTC. Because each stage uses a separate model, teams can swap the underlying LLM, voice model or language support independently, according to MarkTechPost. This modular design allows for customization of the voice agent’s behavior and capabilities.
What should you do?
Developers evaluating voice AI should test the public preview on the Mistral API to assess latency and accuracy for their specific use cases. Security teams should review the data handling practices of any third-party SDKs integrated into their applications.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Mistral AI
- Review the documentation for the LiveKit-based SDK to understand data transmission protocols.
- Test the real-time transcription models against your specific audio environments and accents.
- Monitor the public preview status for potential changes to model availability or pricing.
- Evaluate the context biasing features to improve accuracy for industry-specific terminology.
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 models are included in the Voxtral 3 release?
The release includes three transcription models, one text-to-speech model and a voice agent SDK.
How does Voxtral 3 compare to other speech models?
Small Transcribe 3 ranks third on Artificial Analysis with a 2.1% AA-WER, outperforming Whisper large-v3.
Can the voice agent handle interruptions?
Yes, the agent stops talking when the caller cuts in, with each turn completing in under one second.
