ElevenLabs launches Scribe v3 and v3 Live: 27% lower WER than v2 across 102 languages, 43% lower semantic WER in realtime, and first realtime diarization

Why it matters

【Key points】

· Nov 10, 2026: ElevenLabs launches Scribe v3 for recorded audio and Scribe v3 Live for realtime (model IDs scribe_v3 / scribe_v3_live). Both are in ElevenAPI and are now the default in the ElevenLabs app.

· Accuracy: 27% lower WER than Scribe v2 across 102 languages on an internal version of FLEURS, with gains in every language (German −42%, English −23%, Swahili −34%).

· Realtime: v3 Live cuts semantic WER by 43% and WER by 31% vs v2 Realtime (Greek −77%, Hindi −36%).

· Diarization: v3 makes 12% fewer errors than v2 (17% fewer with 3+ speakers). v3 Live adds realtime diarization for the first time: 3.1% error vs 5.9–7.3% for the other realtime models tested.

• · Also new: context-based turn detection, up to 1,000 keyterms, plain-English formatting rules, and mid-sentence language switching in Live.
• Data residency in the US by default, with EU, India and Singapore options.
• Zero Retention mode and HIPAA eligibility.
• No new pricing was announced.

【Our read】

· Background: after Eleven Realtime speech-to-speech (e/1006), ElevenLabs is completing its voice-agent stack with better listening.

• · Comparison: NVIDIA's open-weight Nemotron 3.5 Transcribe (e/1062) targets similar features.
• Scribe v3 is a closed API that competes on latency, compliance and hosting.
• The benchmarks are ElevenLabs' own (internal FLEURS plus a dataset it says it will open-source soon), averaged per language, so they are not directly comparable to pooled scores.

· Impact: turn detection and realtime diarization address the two most common voice-agent failures, cutting callers off and misattributing speakers. Watch for third-party re-tests once the dataset is released.

【Verifiable】

• ·
• ·