Mistral launches Voxtral 3 speech models
Mistral has released Voxtral 3, a set of transcription and text-to-speech models plus a voice-agent SDK that compete on price and can be self-hosted.
Mistral announced Voxtral 3, a family of speech-to-text and text-to-speech models paired with a voice-agent SDK. The SDK runs listening, reasoning, tool calling and speaking in a single real-time streaming pipeline. Customers can use the models through Mistral's EU-hosted API or run them on their own infrastructure, and choose whether audio is retained.
Transcription. Voxtral Transcribe 3 now covers 50 languages, including all 24 official EU languages and several low-resource Asian and Indic languages. Mistral says the Small variant is state of the art in multilingual transcription and diarization, and that it beats Whisper, ElevenLabs' Scribe v2 and Gemini 3.5 Transcribe. Those results are reported by Mistral and have not been independently checked. The Mini variant has 4B parameters and costs $0.0035 per minute. Both variants support context biasing, which lets users supply key terms to improve accuracy on domain vocabulary without fine-tuning. Mistral did not give a parameter count for Small.
Realtime. Voxtral Realtime Transcribe 3 transcribes with under 500ms of latency in more than 35 languages, with diarization. It also predicts when a speaker has finished a turn, using both acoustic and semantic cues, so agents can reply quickly without cutting users off.
Speech synthesis. Voxtral Mini TTS 3 runs in batch and realtime modes and covers 35+ languages and 50+ dialects. It can be instructed to adjust emotion, pacing and emphasis. At $13 per million characters, Mistral says it is cheaper than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. A "Pro Voice" feature adapts a speaker's voice and style across languages from about 30 minutes of audio.
The announcement does not include full benchmark tables or details on SDK availability.