Pitch Estimation, Voicing Decision, and Noise Spectrum Estimation for Speech Corrupted by High Levels of Additive Noise

This dissertation presents two algorithms that extract parameters which are important to speech processing in high levels of noise. The first algorithm determines whether a signal containing noise corrupted human speech is voiced or not and estimates the fundamental frequency (pitch) of voiced speech. The second algorithm produces an estimate of the additive noise which is corrupting the speech. Previous research related to the voicing decision and pitch estimation has been concentrated at signal-to-noise ratios (SNRs) above 0 dB. Consequently, speech processing requiring the extraction of these parameters in higher levels of noise could not be performed with much success. The research presented in this dissertation concentrates on SNRs around and below 0 dB. Although the algorithm, based on the autocorrelation function, is designed to work well for high levels of noise, good results for the no noise case have been maintained. The idea of a confidence measure for parameter estimation is introduced. Confidence measures are defined and developed for both the voicing decision and the pitch estimation algorithms. Estimation of noise that is corrupting a speech signal has been motivated by the need to enhance the corrupted speech. Previous research has concentrated on speech which is band limited to about 3500 Hz. Therefore, the estimation of the noise corrupting high frequency speech had not been considered. The noise estimation algorithm presented in this dissertation considers the effects of high frequency speech on the noise estimate in addition to the effects of low frequency speech. A new spectral averaging method is introduced which significantly reduces the corrupting effect of the speech components on the noise estimate for SNRs above 0 dB. The algorithm is tested for stationary white noise, stationary non-white noise, and non-stationary white noise