The abstract is read at the publisher; the summary is JClub's.
Three industry speech recognition models trained on hundreds of years of speech outperform humans on 272 speech distortions, yet match human word recognition when linguistic context is removed.
- Why it matters: It was unclear whether modern speech models behave like people in word recognition, or whether they succeed for different reasons.
- What they did: The study measured human and model word recognition across 272 speech distortions, then tested context-removed conditions and varied training data to probe what drives the match.
- The result: Removing linguistic context yields human-level performance, implying the models are superhuman at using statistical dependencies between words, and that training data variability and realism help.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.