Mistral presents Voxtral 3 voice agent SDK and speech models

Mistral presents Voxtral 3 as both a voice agent SDK and a set of speech models.[1] Its broader audio offering covers speech-to-text, text-to-speech and live conversations.[3] The distinction matters for teams evaluating the announcement: the SDK is described as supporting a real-time conversation flow, while the models also address tasks such as transcribing audio and generating speech.[2][3] Mistral says the SDK supports real-time speech in more than 50 languages, but the evidence does not specify separate language counts for each model or task.[1]

Mistral’s documentation describes the Voice Agents SDK as supporting real-time, bidirectional voice conversations.[2] It says the SDK handles microphone capture, speech-to-text, an LLM within the conversation flow, text-to-speech and audio playback.[2] That gives developers a documented path from captured speech through a language-model step to spoken output. Mistral also says its voice agents can answer callers in real time.[3] The documentation does not, in the evidence available here, establish performance levels or describe how the system behaves under particular workloads.

Voxtral 3 models can transcribe both recorded audio files and live audio streams, according to Mistral’s documentation.[3] Transcriptions can include speaker labels and timestamps, and users can apply custom vocabulary.[3] For speech generation, Mistral lists preset voices and cloned voices, with emotion and pace tags available to direct the output.[3] These options give teams specific capabilities to evaluate against their requirements, but the evidence does not report accuracy figures, voice-cloning conditions or comparative results.

Mistral says Voxtral 3 can be hosted in the EU and can be self-deployed.[1] Those options are relevant for teams weighing where and how to run a voice system, though the evidence does not provide pricing, deployment requirements or details about data handling. Mistral’s documentation also says users can send audio to a chat model to summarize a recording or ask questions about it.[3] Separately, LiveKit’s documentation includes a guide to using a Mistral text-to-speech plugin with LiveKit Agents.[4] That establishes a documented plugin path, not that the Voxtral 3 SDK itself is built on LiveKit.