Speech/silence discrimination of noisy speech

Abstract

In this thesis, theoretical and experimental investigation for detecting the presence of speech in noisy speech that is corrupted by white Gaussian noise has been done. In one algorithm, speech/silence discrimination is made based on the energy and autocorrelation of input signal. The threshold of energy is adapted by comparing the current threshold values and the energy and autocorrelation of the incoming noisy speech. Computer simulation has been done with various real speech, clean speech and noisy speech with signal-to-noise ratios (SNR``s) of 20, 10 and 0 dB. The percentage of discrimination errors were less than 6 percents. In another algorithm we have done preprocessing by Wiener filtering the input signal and used the same algorithm described above. The noisy input speech is first Weiner filtered to reduce the effect of noise. Then, speech/silence decision of the filtered signal is done by measuring its energy and autocorrelation. According to our computer simulation results, the percentage of discrimination errors decreases when the input noisy signal is preprocessed in the case that the input signal is extremely noisy. In addition, design of the new speech detector hardware has been considered.