NatWest benchmarks British-accent ASR: Cohere Transcribe hits 15.3% WER
Nils_Reimers · x · 2026-10-21
Independent research from NatWest Group, published as the arXiv paper "Mind the Accent Gap" with the new CavaBench benchmark of spoken financial queries, evaluates ASR models on British accents. Cohere Transcribe achieves a 15.3% word error rate, beating Whisper Turbo and Parakeet with one of the smallest cross-accent gaps—without fine-tuning. Key findings: WER strongly predicts downstream tool-calling accuracy (r=-0.93), yet can mask task-level failures that vary substantially across models and acoustic conditions. Directly relevant to voice-agent model selection in finance.
More from Models
- 35B MoE at 20.5 tok/s on an iPhone: OpenBMB's Lookahead Expert Scheduling skips no experts — GasSmooth7439 · 2026-10-21
- ICA Lens in action: decomposing Qwen residuals reveals clean city/France features — furongh · 2026-10-21
- $200 subscription covers $655 of API usage — 3900% ROI on Claude Code/Codex — umeshonai · 2026-10-21
- NVIDIA Kumo Forecast Pairs Pretrained Time-Series Representations with Pattern Retrieval — Hugging Face Blog · 2026-10-21
- Reddit user suspects Gemini 6.1 Ultra usage limits quietly shrunk, seeks independent benchmarks — Reiszecke · 2026-10-21
- ChatGPT's 'memory full' limit hits users after 4-15 messages, even paid tiers — CommercialLynx1791 · 2026-10-21