Hierarchical spectro-temporal features for robust speech recognition

This thesis proposes to use Hierarchical Spectro-Temporal (HIST) features to improve the robustness of automatic speech recognition systems. The presented HIST features are inspired by recent neurobiological knowledge about the auditory cortex and are complementary to conventional speech features. The spectro-temporal feature extraction is hierarchically organized: a fi rst layer detects local and simple spectro-temporal patterns, and a second one combines them into complex features. The shapes of the spectro-temporal receptive fi elds, at the different steps of the hierarchy are learned in an unsupervised manner. The concept of hierarchical spectro-temporal processing is fi rst validated by adapting a biologically-inspired visual object recognition system to a words recognition task. Afterwards, the performance of the HIST features is tested on a continuous spoken digits recognition task in noisy conditions and shows that the features are complementary to the state-of-the-art speech features