Mistral launches Voxtral 3 with a voice agent SDK
The SDK lets a voice agent call tools while listening and speaking, with turns kept under one second and speech that stops when a caller interrupts.
The LiveKit-based pipeline runs transcription, reasoning, tool use, and speech generation concurrently. Teams can swap the language model, voice, or language without rebuilding the agent. Transcription supports more than 50 languages; speech generation covers more than 35 languages and over 50 dialects. The stack is available through Mistral’s APIs or private deployments, where customers can keep audio inside their own environment and control conversation retention.
TestingCatalog's note
The LiveKit-based pipeline runs transcription, reasoning, tool use, and speech generation concurrently. Teams can swap the language model, voice, or language without rebuilding the agent. Transcription supports more than 50 languages; speech generation covers more than 35 languages and over 50 dialects. The stack is available through Mistral’s APIs or private deployments, where customers can keep audio inside their own environment and control conversation retention.
TestingCatalog's note
score 6