Mistral Voxtral 3: 2.1% AA-WER ASR and a LiveKit Voice Agent SDK

Mistral AI shipped Voxtral 3 on October 20, 2026: three speech-to-text models, one text-to-speech model, and a LiveKit-based voice agent SDK, with all four models in public preview on the Mistral API. MarkTechPost reports each conversational turn completes in under one second.

Model family and SDK architecture

Voxtral Small Transcribe 3 is the flagship batch transcription and diarization model; Mistral's docs also list translation, speech Q&A, summarization, and function calling. Voxtral Mini Transcribe 3 is disclosed at 4B parameters and targets lower-cost batch jobs. Voxtral Mini Transcribe Realtime 3 runs streaming ASR over WebSocket with end-of-turn detection. Voxtral Speech 3—called Voxtral Mini TTS 3 in MarkTechPost's reporting—is the TTS model. All transcription models support diarization and context biasing, which accepts key terms or free text to improve recall on domain vocabulary. Batch transcription covers 50 languages, including 24 official EU languages.

The LiveKit SDK chains the stages into one streaming pipeline: Realtime Transcribe 3 listens and detects end of turn, a pluggable LLM such as Mistral Small 4 handles reasoning and tool calls, and Voxtral Speech 3 generates the reply. Because each stage is a separate model, teams can swap the LLM, voice, or language without restructuring the pipeline. The agent handles SIP phone calls and WebRTC browser/app sessions, and interruption detection stops it when the caller cuts in.

Benchmark performance

Accuracy figures carry vendor attribution. On FLEURS 45 with forced language, Mistral reports Small Transcribe 3 at 4.1% WER and Mini at 4.6%; MAI-Transcribe-2 ties Small at 4.1%, while Whisper large-v3 posts 16.0%. On five public diarization sets, Mistral records Small at 18.4% cpWER and Mini at 18.6%, against 19.0% for Pyannote Precision 3 and 38.4% for Gemini 3.5 Transcribe.

Artificial Analysis independently measures Small at 2.1% AA-WER, third behind Fun-Realtime-ASR-preview at 1.7% and MAI-Transcribe-2 at 2.0%. The speed gap is large: Small's speed factor is 65.8x, against 307.5x for MAI-Transcribe-2.

Realtime accuracy is not Voxtral's strongest tier. On FLEURS 35, Mini Transcribe Realtime 3 posts 7.4% WER at 240 ms, 6.4% at 480 ms, and 5.7% at 2,400 ms. Gemini 3.5 Realtime Transcribe scores 4.2%, ahead of every Voxtral setting, while MAI-Transcribe-2 Streaming scores 5.8%, ahead of the two lower-latency Voxtral settings but behind the 2,400 ms setting. GPT Live Transcribe posts 8.4% and Scribe v2 Realtime 8.7%, trailing all three Voxtral settings. MarkTechPost reports these from Mistral's self-reported benchmark disclosures.

Pricing and availability

Per Mistral's price table, Mini Transcribe 3 batch costs $0.0035/min, Small Transcribe 3 $0.004/min, and Mini Transcribe Realtime 3 $0.008/min. The voice agent API runs at $0.04/min with Mistral Small 4 and $0.08/min with any other reasoning model. Voxtral Speech 3 is $13 per 1M characters; Mistral says that undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. The TTS model streams with roughly 90 ms time-to-first-audio, exposes 34 emotion tags and 2 pace tags, supports zero-shot voice cloning, and covers 9 languages. A Pro Voice mode adapts a speaker's style from 30 minutes of audio.

All models run on Mistral's EU, US, and global endpoints. MarkTechPost reports OpenRouter lists all three new transcription models with EU-hosted, zero-data-retention endpoints. On-premises deployment is available through Mistral's sales channel; GPU and VRAM requirements are not disclosed.

AI Mastery analysis

Three architectural points matter. First, the realtime gap—7.4% WER at 240 ms versus Gemini 3.5 Realtime's 4.2%—is a real limitation for sub-300 ms phone-tree and support agents, though context biasing can recover accuracy on predictable term sets. Second, the throughput gap matters more in batch pipelines: Small's 65.8× Artificial Analysis speed factor looks weak against MAI-Transcribe-2's 307.5× even though $0.004/min keeps 10,000 minutes at $40. Third, the SDK's swappable LLM stage aligns with the infrastructure-not-weights pattern, but a slower external reasoning model could push coordination latency beyond the reported sub-one-second turn.

No open weights were announced, so self-hosted teams depend on Mistral's on-premises sales track rather than a community runtime. Voxtral 3 is strongest today as a diarization-heavy, batch-first voice stack with third-party accuracy data; its value beyond that depends on closing the realtime accuracy gap.

Sources

• Primary: Voxtral TTS 3 - Mistral AI | Mistral Docs
• Reported by: MarkTechPost: Mistral AI Releases Voxtral 3: Speech-to-Text, TTS and a Voice Agent SDK