Extraction of Spectro-Temporal Speech Cues for Robust Automatic Speech Recognition

摘要

This work analyzes the use of spectro-temporal signal characteristics with the aim of improving the robustness of automatic speech recognition (ASR) systems. Experiments that aim at the robustness against extrinsic sources of variability (such as additive noise) as well as intrinsic variation of speech (changes in speaking rate, style, and effort) are presented. Results are compared to scores for the most common features in ASR (mel-frequency cepstral coefficients and perceptual linear prediction features), which account for the spectral properties of short-time segments of speech, but mostly neglect temporal or spectro-temporal cues. Intrinsic variations were found to severely degrade the overall ASR performance. The performance of the two most common feature types was degraded in much the same way, whereas the proposed spectro-temporal features exhibit a different sensitivity against intrinsic variations, which suggests that classic and spectro-temporal feature types carry complementary information. Furthermore, spectro-temporal features were shown to be more robust than the baseline system in the presence of additive noise.