Mistral AI launches Voxtral 3 for real-time voice agents

Mistral AI has introduced Voxtral 3, a family of speech recognition and speech generation models designed for real-time voice agents. In an official blog post, Mistral AI announces that the release also includes an SDK that combines transcription, reasoning, tool use, and spoken responses in one streaming system.

The company is targeting uses such as contact centers, help desks, virtual assistants, and hands-free workplace copilots. The system can connect to company data and business tools, allowing an agent to verify an identity, check a bill, schedule an appointment, or retrieve information during a live call.

Transcription, speech and tools in one pipeline

Voxtral Small Transcribe 3 supports more than 50 languages and can identify different speakers in a conversation. It also supports context biasing, which lets organizations provide product names, technical terms, or other relevant vocabulary to improve recognition without retraining the model.

For live applications, Mistral offers Voxtral Mini Transcribe Realtime 3. The company says it produces transcription in under 500 milliseconds in more than 35 languages. It can also detect when a person has finished speaking, a feature intended to reduce awkward pauses and prevent an agent from interrupting too early.

Voxtral Mini TTS 3 generates speech in more than 35 languages and 50 dialects. According to Mistral, it adjusts pace, emphasis, and emotional delivery from the surrounding text. The company also offers voice cloning and a Pro Voice option that can adapt to a speaker’s style from 30 minutes of audio.

The SDK is built on LiveKit and allows developers to swap the reasoning model, voice, or language separately. Mistral says the full system can complete a conversational turn in less than one second.

Voxtral 3 is available through Mistral’s API and for self-hosted deployments. Pricing starts at $0.0035 per minute for batch transcription, while the voice agent API starts at $0.04 per minute when using Mistral Small 4 for reasoning.