Mistral AI releases Voxtral 3 speech models and voice agent SDK
Mistral AI has released Voxtral 3, a family of three transcription models, one text-to-speech model and a LiveKit-based voice agent SDK. All four models are in public preview on the Mistral API.
Voxtral 3 includes Voxtral Small Transcribe 3 for batch transcription and diarization, Voxtral Mini Transcribe 3 for cheaper and faster batch jobs, Voxtral Mini Transcribe Realtime 3 for streaming ASR over WebSocket with end-of-turn detection, and Voxtral Speech 3 for text-to-speech. Mistral lists the Mini transcription model at 4 billion parameters and does not disclose the size of Small Transcribe 3. Both batch models list a 32,000-token context. The company did not disclose GPU or VRAM requirements. Mistral says Mini is suited to resource-constrained environments.
All transcription models support diarization and context biasing, which takes key terms or free text to improve recall, Mistral said. Batch transcription covers 50 languages, including 24 official EU languages. The models can run on Mistral's European, US and global endpoints, or on customer infrastructure through Mistral sales. OpenRouter also lists all three new models with EU-hosted, zero-data-retention endpoints.
On performance, Mistral said Small Transcribe 3 scores 2.1% AA-WER and ranks third on Artificial Analysis, while Mini scores 2.2%. Only Fun-Realtime-ASR-preview at 1.7% and MAI-Transcribe-2 at 2.0% rank higher. Mistral's speed factor for Small is 65.8x, compared with 307.5x for MAI-Transcribe-2. The realtime model trades latency for accuracy on FLEURS 35: 5.7% WER at 2,400 ms, 6.4% at 480 ms and 7.4% at 240 ms. Gemini 3.5 Realtime Transcribe scores 4.2% and MAI-Transcribe-2 Streaming 5.8%, while GPT Live Transcribe and Scribe v2 Realtime trail every Voxtral setting at 8.4% and 8.7%, according to the report.
Mistral-reported diarization cpWER was 18.4% for Small and 18.6% for Mini across five public sets, ahead of Pyannote at 19.0% and MAI-Transcribe-2 at 20.3%. On FLEURS 45, Mistral reported WER of 4.1%/4.0% for Small with forced/inferred language and 4.6%/4.5% for Mini; Whisper large-v3 was 16.0%/14.9%. Keyphrase recall on Earnings 22 was 94.6% for Small with context and 81.4% without.
The voice agent SDK runs three stages as one streaming pipeline. Realtime Transcribe 3 listens and detects when the caller stops speaking; a general LLM such as Mistral Small 4 reasons and calls tools; Voxtral Speech 3 replies in the tone of the conversation so far. Mistral says each turn completes in under one second and the agent stops talking when the caller cuts in. It handles phone calls over SIP and browser or app sessions over WebRTC. Because each stage is a separate model, teams can swap the LLM, voice or language.
Voxtral Speech 3 costs $13 per 1 million characters. Mistral says that undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. The model page lists about 90 ms time-to-first-audio and 34 emotion tags, with explicit controls for emotion, pacing and emphasis. Pro Voice adapts a speaker's style from 30 minutes of audio. Artificial Analysis places it in the top 12 of both TTS arenas, according to the report.
Prices per minute are $0.0035 for Mini Transcribe 3 batch, $0.004 for Small Transcribe 3 and $0.008 for Mini Transcribe Realtime 3. The voice agent API costs $0.04 per minute with Mistral Small 4 and $0.08 per minute with any other reasoning model. Voxtral Speech 3 is priced at $13 per 1 million characters for batch and realtime.