Estimating Reliability of Speech Recognition Results Using Recognition Results Change Degree When Minute Changes to Speech Waveform

Abstract

Speech recognition technology has made remarkable progress in recent years, but 100% of speech recognition results are not guaranteed to be recognized due to the nature of the speech recognition mechanism. From the standpoint of developing applications that incorporate such speech recognition software, it is necessary to know how many misrecognition errors are included in the recognition results for the speech input as this will affect the application’s performance. In this paper, we propose a method to estimate a reliability of speech recognition results by observing how they change when slight modifications are intentionally applied to the input waveform. Specifically, the two types of minute change are processed to waveform. One is to slightly shift the start points of speech parts. Another is to add slight white noise to the speech waveform. By analyzing recognition results before and after these modifications, we find a correlation between the recognition results changing degrees and the words error rate (WER). This suggests that the recognition result change rate is an effective index for estimating the reliability of output speech recognition results.