Speech recognition by man and machine

While human listeners have little problems in dealing with the strong variation in spoken language, the same cannot be said about automatic speech recognition (ASR). This work compares recognition performance of man and machine with the aim of learning from the distinct errors between these two. Based on the differences, the signal processing mechanisms are analyzed that are suitable to increase the robustness of ASR. The comparison focuses on the influence of intrinsic variation of speech, i.e., changes in speaking rate, effort and style, as well as dialect and accent. The outcome of the experiments suggests that the processing of temporal cues in ASR bears room for improvement. Therefore, spectro-temporal features are employed as input to ASR systems, which results in an increase of recognition performance for varying speaking effort and speaking style compared to standard features. This documents the usefulness of spectro-temporal and temporal information for automatic recognizers.

Auteur

Bernd T. Meyer studied physics at the University in Oldenburg, and received his diploma and Ph.D. in 2004 and 2009, respectively. He has been working on the improvement of automatic speech recognizers and modeling human speech perception both in Oldenburg and the International Computer Science Institute in Berkeley, CA.