ElevenLabs Scribe v3 Live tops STT leaderboard with 1.89% WER across 21 models
lukeharries · x · 2026-11-11
Cekura's live speech-to-text benchmark puts ElevenLabs Scribe v3 Live at #1 of 21 models with a 1.89% overall word error rate, and 2.05% on real phone call recordings (Ocular-licensed data).
- Reson8 Rumble-2 is the speed leader: 18ms final-text P50 (vs 57ms for Scribe v3 Live), 2.05% WER, and far cheaper at $0.12/hour vs Scribe's listed $0.39/hour (Scribe v3 Live pricing not yet published).
- Fish Audio Transcribe Stream 1 ranks third at 2.21% WER.
Takeaway: frontier voice models are shipping fast — Rumble-2 wins latency, Scribe v3 Live wins accuracy.
More from Multimodal
- MiniMax H3 Image nails intricate exploded views and accurate Japanese text in demo — Hailuo_AI · 2026-11-11
- MiniMax H3 Image local test yields Japanese-film-grade images, fast and cool-running — Hailuo_AI · 2026-11-11
- One prompt plus a PDF: Fable 5.2 generates narrated video of a Tuscan ruin restored — FinanceYF5 · 2026-11-11
- Ex-film composer builds dream ambient music workstation with Fable 5.2 in one week — FinanceYF5 · 2026-11-11
- AI-generated 2001: A Space Odyssey ape scene stuns Reddit with film-quality visuals — Megalania34 · 2026-11-11
- Fable 5.2 builds 3D animation from scratch, closing the faceless YouTube skill gap — FinanceYF5 · 2026-11-11