Auditory modeling for robust speech recognition in noisy environments

Abstract

In this dissertation, a computationally efficient auditory model, zero-crossings with peak amplitudes (ZCPA), motivated by mammalian auditory periphery is developed to extract reliable features from speech signals even in noisy conditions. Compared with other auditory models, the developed auditory model is computationally efficient and free from many unknown parameters. The noise-robustness of the developed model is shown analytically as well as experimentally. Speaker-independent isolated word recognition experiments demonstrate that the developed auditory model outperforms other feature extraction methods especially at low signal-to-noise ratio (SNR) conditions corrupted by not only white Gaussian noise but also several real-world noises. Improvements in the recognition rates are more eminent at very low SNR conditions. Detail frequency responses of the filterbank and microphone gains are not critical. Both spectral and cepstral representations of the model outputs are considered, and the cepstral representation shows improved recognition accuracy with less number of coefficients than the spectral representation. Also, several different lengths of time have been tried to obtain good time-derivative features of the developed auditory model.