Obtaining a Spectral Estimate of Time Varying, Broadband Noise Corrupting a Speech Signal

In this thesis an algorithm is presented which provides an estimate of the noise magnitude spectrum present with a noise corrupted speech signal. Speech parameters such as the fundamental frequency, formant frequencies, and voiced/unvoiced categorization are known to the algorithm; however, these parameters need not be known precisely. No prior information as to the spectral density of the noise present need be known. It only need meet the criteria of broadband and slowly time varying. The algorithm is analyzed with white, broadband, and time varying noise sources. The algorithm is shown to be successful, especially for signal-to-noise ratios less than 10dB. In addition, accurate estimates for modes of noise with bandwidths as small as 200Hz are shown. Convergence times for time varying noise are also discussed. Given the results and the stated restrictions on the noise, the algorithm is able to obtain an accurate estimate of the noise magnitude spectrum present with a noise corrupted speech signal