An event-related brain potential study on the impact of speech recognition errors

Abstract

Most automatic speech recognition (ASR) systems, which aim for perfect transcription of utterances, are trained and tuned by minimizing the word error rate (WER). In this framework, even though the impact of all errors is not the same, all errors (substitutions, deletions, insertions) from any words are treated in a uniform manner. The size of the impact and exactly what the differences are remain unknown. Several studies have proposed possible alternatives to the WER metric. But no analysis has investigated how the human brain processes language and perceives the effect of mistaken output by ASR systems. In this research we utilize event-related brain potential (ERP) studies and directly analyze the brain activities on the impact of ASR errors. Our results reveal that the peak amplitudes of the positive shift after the substitution and deletion violations are much bigger than the insertion violations. This finding indicates that humans perceived each error differently based on its impact of the whole sentence. To investigate the effect of this study, we formulated a new weighted word error rate metric based on the ERP results: ERP-WWER. We re-evaluated the ASR performance using the new ERP-WWER metric and compared and discussed the results with the standard WER.