EARLY AUDITORY PROCESSING INSPIRED FEATURES FORROBUST AUTOMATIC SPEECH RECOGNITION

摘要

In this paper, we derive bio-inspired features for automaticspeech recognition based on the early processing stages inthe human auditory system. The utility and robustness ofthe derived features are validated in a speech recognitiontask under a variety of noise conditions. First, we developan auditory based feature by replacing the filterbank analysisstage of Mel-frequency cepstral coefficients (MFCC)feature extraction with an auditory model that consists ofcochlear filtering, inner hair cell, and lateral inhibitory networkstages. Then, we propose a new feature set that retainsonly the cochlear channel outputs that are more likely to firethe neurons in the central auditory system. This feature set isextracted by principal component analysis (PCA) of nonlinearlycompressed early auditory spectrum. When evaluatedin a connected digit recognition task using the Aurora 2.0database, the proposed feature set has 40% and 18% averageword error rate improvement relative to the MFCC andRelAtive SpecTrAl (RASTA) features, respectively.