Mistral AI has unveiled Voxtral 3, a family of models built for voice agents that talk with people in real time. The package combines listening, reasoning, use of external tools and spoken replies in a single streaming pipeline, meaning each stage starts without waiting for the previous one to finish.
At its core are three components. The first is the speech-to-text model Voxtral Small Transcribe 3, which converts speech into text in more than 50 languages. Alongside it sits Voxtral Mini Transcribe 3, while the Voxtral Mini Transcribe Realtime 3 variant is built for live listening, with a claimed latency under 500 milliseconds. The second is the text-to-speech model Voxtral Speech 3, which speaks in more than 35 languages and 50 dialects and adapts tone and emotion to the conversation. The third is an SDK, the toolkit developers use to build applications, based on LiveKit, an open-source platform for real-time audio and video communication.
GitHub - khalooei/Voxtral-AI-Demo-Local-Interface: Voxtral is a state...
The SDK links these pieces to a general language model, such as Mistral Small, which reasons and decides when to call an external service. An agent can thus book an appointment or check an invoice while the caller is still on the line. Mistral points to contact centers, service desks and virtual assistants as use cases.
On pricing, transcription starts at $0.0035 per minute. The package can be used through Mistral's APIs, with endpoints hosted in the European Union, the United States and globally, or installed on the customer's own infrastructure, so audio data stays under its control. That matters in Europe, where voice conversations are often subject to strict privacy rules.
GitHub - Tamles/voxtral-webgpu · GitHub
The performance figures Mistral reports are compared with Whisper, Gemini and GPT models on multilingual transcription, speaker identification and pronunciation. These are the company's own numbers and have not yet been checked by independent tests.
The launch places Mistral in a crowded market, where OpenAI, ElevenLabs, Deepgram and Google compete for voice agents. The difference Mistral wants to stress is price combined with the option of keeping data in-house. For European companies seeking an alternative to US services, the local deployment option could be the deciding factor.
No comments yet. Be the first!