Voxtral
A speech interaction model with real-time speaker diarization and end-of-turn detection capabilities, supporting two deployment modes: self-hosted infrastructure or official managed hosting.
Timeline (last 90 days)
Currently the public information is only a single tweet, lacking specific performance metrics, latency data, and benchmark results, and Voxtral's actual performance still awaits more detailed technical disclosure.
Voxtral demonstrated real-time speaker diarization capability.
Voxtral demonstrated real-time end-of-turn detection capability.
Voxtral supports deployment on users' own infrastructure or use of the official cloud service.
Mistral AI's product suffixes encode function rather than scale: Pixtral corresponds to vision, Codestral to code, Mathstral to math, Devstral to agentic coding, and Voxtral to speech.
Voxtral uses a 3-billion-parameter text model as its backbone and converts audio into continuous embeddings via an audio encoder that are fed directly into the decoder, which differs from Whisper's approach of feeding into the decoder via cross-attention.
Voxtral's audio encoder produces one audio token every 80 milliseconds, i.e., a fixed frame rate of 12.5 tokens per second.
Users can load audio models such as Voxtral and Qwen2-Audio via FastModel.
All Facts (8)
Currently the public information is only a single tweet, lacking specific performance metrics, latency data, and benchmark results, and Voxtral's actual performance still awaits more detailed technical disclosure.
50%UnverifiedVoxtral demonstrated real-time speaker diarization capability.
50%UnverifiedVoxtral demonstrated real-time end-of-turn detection capability.
50%UnverifiedVoxtral supports deployment on users' own infrastructure or use of the official cloud service.
50%UnverifiedMistral AI's product suffixes encode function rather than scale: Pixtral corresponds to vision, Codestral to code, Mathstral to math, Devstral to agentic coding, and Voxtral to speech.
50%UnverifiedVoxtral uses a 3-billion-parameter text model as its backbone and converts audio into continuous embeddings via an audio encoder that are fed directly into the decoder, which differs from Whisper's approach of feeding into the decoder via cross-attention.
50%UnverifiedVoxtral's audio encoder produces one audio token every 80 milliseconds, i.e., a fixed frame rate of 12.5 tokens per second.
50%UnverifiedUsers can load audio models such as Voxtral and Qwen2-Audio via FastModel.
50%