Speech Enhancement In The Fan-chirp Transform Domain

Performance of speech processing system will be degraded under noise, whichimposes a significant consequence on speech quality. Thus speech enhancement is of greatnecessity for noise rejection as a front-end processing of speech systems. This thesisstudies speech enhancement in the Fan-chirp transform (FChT) domain, which is the firstapplication of FChT into the speech enhancement domain. Based on the feature ofFan-chirp spectra, we propose an estimation method of the chirp rate as well as the noiseestimation algorithm in the FChT domain. By combining a frequency-domain comb filterwith the minimum mean-square error estimator based on the super-Gaussian mixturemodel, speech enhancement is implemented in the FChT domain. The main works areexplained as follows:Firstly, we propose an estimation method for the chirp rate, which is a key parameterof FChT. The chirp rate can be obtained by nearly exhaustive search of the maximum ofthree proposed measures, and the estimates are accurate even in low signal-noise-ratio.Secondly, we present a noise estimation algorithm in the FChT domain, whichbenefits from the high concentration of Fan-chirp spectra. Comparisons with two otherclassic algorithms of noise estimation show that the proposed noise estimation algorithmcan better track the noise spectra.Finally, the probability distribution model of Fan-chirp spectra has been studied andthe noisy Fan-chirp spectra have been better described with the super-Gaussian model. Thespeech is enhanced in the FchT domain by combining a frequency-domain comb filter andthe gain of minimum mean-square error estimator based on the super-Gaussian mixturemodel. Experiments are conducted in both white and babble noise environments withdifferent signal-noise-ratios. Compared with two other classic algorithms, the proposedalgorithm shows its validity for subjective and objective criteria.