Supervised and unsupervised feature extraction from a cochlearmodel for speech recognition

摘要

The authors explore the application of a novel classificationnmethod that combines supervised and unsupervised training, and comparenits performance to various more classical methods. The authors firstnconstruct a detailed high dimensional representation of the speechnsignal using Lyon's cochlear model and then optimally reduce itsndimensionality. The resulting low dimensional projection retains theninformation needed for robust speech recognition