Robust speech recognition based on spectro-temporal processing

In this thesis, novelle spectro-temporal feature extraction techniques are evaluated for enhancing the robustness of automatic speech recognition systems (ASR) in adverse acoustical conditions. Recent physiological and psychoacoustical findings indicate that spectro-temporal processing plays an important role in human speech perception. Therefore, sigma-pi cells and Gabor filter functions are investigated as secondary feature extraction methods based on a spectro-temporal representation. Especially the Gabor features are versatile enough to include cepstral features and purely temporal filtering as special cases, while additionally aiming at combined spectro-temporal modulations. A data driven feature selection method is applied for feature set optimization. For small vocabularies, both types of features are shown to increase the robustness of ASR systems. Sigma-pi cells also allow for estimating the speech-to-noise ratio of an input signal solely based on low spectro-temporal modulation. The Gabor based Tandem feature sets increase the performance of the Qualcomm-ICSI-OGI system for the Aurora task, when concatenating the two streams. engl.