Early auditory processing inspired features for robust automatic speech recognition

Abstract

In this paper, we derive bio-inspired features for automatic speech recognition based on the early processing stages in the human auditory system. The utility and robustness of the derived features are validated in a speech recognition task under a variety of noise conditions. First, we develop an auditory based feature by replacing the filterbank analysis stage of Mel-frequency cepstral coefficients (MFCC) feature extraction with an auditory model that consists of cochlear filtering, inner hair cell, and lateral inhibitory network stages. Then, we propose a new feature set that retains only the cochlear channel outputs that are more likely to fire the neurons in the central auditory system. This feature set is extracted by principal component analysis (PCA) of nonlinearly compressed early auditory spectrum. When evaluated in a connected digit recognition task using the Aurora 2.0 database, the proposed feature set has 40% and 18% average word error rate improvement relative to the MFCC and RelAtive SpecTrAl (RASTA) features, respectively.