Mistral AI has unveiled Voxtral 3, a complete speech model family paired with a voice agent SDK that takes an AI assistant from listening to reasoning to speaking in a single streaming pipeline.
The release ships with three transcription models, one text-to-speech model, and an SDK built on LiveKit. All four models are currently in public preview on the Mistral API, with Mistral API endpoints covering EU, US, and global regions.
The standout is Voxtral Small Transcribe 3, which scores 2.1% AA-WER on Artificial Analysis — ranking #3 among speech-to-text models, trailing only Microsoft’s MAI-Transcribe-2 and Fun-Realtime-ASR-preview. The flagship batch model also leads on diarization, posting an 18.4% cpWER across five public benchmarks, beating Pyannote’s 19.0%.
Voxtral 3 speech models (Credits: Mistral AI)
A Family of Four Models
The lineup covers every speech need. Voxtral Mini Transcribe 3 is a 4B-parameter model aimed at cheaper, faster batch jobs in resource-constrained environments. Voxtral Mini Transcribe Realtime 3 handles streaming speech recognition over WebSocket with end-of-turn detection. And Voxtral Speech 3 — also called Mini TTS 3 — powers the voice output.
All transcription models support diarization and context biasing, letting users feed in key terms so jargon and names are transcribed correctly. Batch transcription covers 50 languages, including all 24 official EU languages. On the FLEURS 45 benchmark, Small Transcribe 3 posts a 4.0–4.1% word error rate — far ahead of Whisper large-v3’s 14.9–16.0%.
The Voice Agent SDK
The LiveKit-based SDK chains three stages into one pipeline: Realtime Transcribe 3 listens and detects when the caller stops speaking, a general LLM such as Mistral Small 4 reasons and calls tools, and Voxtral Speech 3 replies in a tone matching the conversation. Each turn completes in under one second, and the agent stops talking when the caller interrupts.
The SDK handles phone calls over SIP and browser or app sessions over WebRTC. Because every stage is a separate model, teams can swap the LLM, voice, or language independently.
On pricing, Voxtral Speech 3 costs $13 per 1 million characters, which Mistral says undercuts OpenAI’s TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. The TTS model offers explicit controls for emotion, pacing, and emphasis, with 34 emotion tags and ~90ms time-to-first-audio. A Pro Voice mode can adapt a speaker’s style from just 30 minutes of audio.
Also Read: ChatGPT Now Lets You Lock Your Chats Behind a PIN and Anthropic’s Cloud Partnerships Are Getting Messy As AWS, Google And Microsoft See It As A Rival
Source: MarkTechPost — read the full story there.
