Electronic International Standard Serial Number (EISSN)
In this paper, we present advances in the modeling of the masking behavior of the human auditory system (HAS) to enhance the robustness of the feature extraction stage in automatic speech recognition (ASR). The solution adopted is based on a nonlinear filtering of a spectro-temporal representation applied simultaneously to both frequency and time domains-as if it were an image-using mathematical morphology operations. A particularly important component of this architecture is the so-called structuring element (SE) that in the present contribution is designed as a single three-dimensional pattern using physiological facts, in such a way that closely resembles the masking phenomena taking place in the cochlea. A proper choice of spectro-temporal representation lends validity to the model throughout the whole frequency spectrum and intensity spans assuming the variability of the masking properties of the HAS in these two domains. The best results were achieved with the representation introduced as part of the power normalized cepstral coefficients (PNCC) together with a spectral subtraction step. This method has been tested on Aurora 2, Wall Street Journal and ISOLET databases including both classical hidden Markov model (HMM) and hybrid artificial neural networks (ANN)-HMM back-ends. In these, the proposed front-end analysis provides substantial and significant improvements compared to baseline techniques: up to 39.5% relative improvement compared to MFCC, and 18.7% compared to PNCC in the Aurora 2 database.