Modern speech recognition systems exhibit human-like word recognition but superhuman use of statistical dependencies between words
In the authors' words
Advances in machine learning have yielded machine systems that approach human performance on real-world perception tasks. For speech recognition, it remains unclear whether such models' behavior resembles that of humans. We addressed this question by measuring word recognition abilities in humans and models across 272 speech distortions, providing a detailed behavioral fingerprint. Three recent industry speech recognition models trained on hundreds of years of speech outperformed humans. However, in conditions where linguistic context was removed, they matched human performance well, suggesting that the models are superhuman in their ability to leverage statistical dependencies between words. Experiments manipulating model training data suggest that training data variability and realism contribute to the match to human perception. Overall, the results show that superhuman performance is now possible with systems that are optimized over enough data and suggest that accurate models of human speech recognition require optimization over appropriate tasks and amounts of speech.
Appeared: Thursday, November 19. bioRxiv. Preprint, not yet peer-reviewed.