Spectral Analysis of Noise Corrupted Speech using Auditory Models

Model based spectral estimation methods have become popular because of their noise robustness and high resolution. These are useful properties in many signal processing applications such as speech recognition and speech enhancement in noise. However, in speech processing, the performance of the human auditory system is far superior to most existing techniques in terms of noise robustness and acuteness. This thesis presents an analysis of a novel auditory model of the inner ear to achieve robust spectral estimation. To obtain a better estimate of the spectral magnitude, a new method for determining aggregate synchrony is developed without being constrained to resemble the human auditory system. Quantitatively, the aggregate synchrony spectrum is able to resolve sinusoids as close as l00Hz (.01) even below - 5dB SNR. Using mathematical analysis the first stage of the model, bandpass filtering, is shown to directly determine spectral resolution and noise robustness. Statistical evaluation results are presented for the performance factors of bias, variance, resolution and noise threshold of the auditory spectral estimator. Also, natural speech signals are used to evaluate the performance of the new spectral estimator