The Study On Speech Enhancement Algorithm Based On CS Theory

Speech signal is a kind of non-stationary and time-varying signal. Voice is the mostimportant and most common method for information exchange of human being. Speech signalprocessing, which mainly includes speech coding, speech synthesis, speech recognition andspeech enhancement, has a very wide range of application. Usually, the researchers conductvarious speech signals processing in relatively pure conditions. But speech signal alwayscontaminated by various noises in real life environment. The noise will seriously affect thequality and the intelligibility of speech signal. This thesis reviews the advantages anddisadvantages of speech enhancement algorithms by focusing on the classical spectralsubtraction algorithm and sub-space based speech enhancement algorithm. A novel speechenhancement algorithm is proposed by introducing the compressed sensing (CS) theory. Themain works and novelties are as follows:Firstly, this thesis proposes a novel speech endpoint detection method by calculating thequotient of short-time zero-crossing and short-time energy of speech signal. A threshold isobtained according to the quotient of the start segment of speech signal. Then each frame ofspeech signal is classified into non-speech segment or speech segment. The speechsegments are used for speech enhancement. We conduct an endpoint detection experimenton the pure speech signal and the noisy speech signal contaminated by train noise with0dB ofSignal-to-Noise Rate (SNR). The short-time energy, the short-time zero-crossing, theshort-time entropy of spectrum, the product of short-time energy and short-time zero-crossingand the quotient of short-time zero-crossing and short-time energy. Experimental resultsdemonstrate that the quotient of short-time zero-crossing and short-time energy is more robustfor endpoint detection on noisy speech.Secondly, the thesis proposes a novel approach for speech enhancement based oncompressed sensing theory according to the differences of sparcity between the noise signaland the speech signal in the discrete cosine transform (DCT) domain. We scarify each frameof noisy speech by the DCT. The partial Hadamard ensemble is employed as a sensing matrixto achieve compressive measurement of DCT coefficients. Speech signal is thenre-constructed with a modified orthogonal matching pursuit (OMP) algorithm by using twothresholds to realize the speech enhancement. The convergence should be slow on theoriginal OMP algorithm when noise signal contain speech like components. In the thesis,the iteration procedure of the OMP algorithm is controlled by introducing two energythresholds according to the similarity between the residual signal and the column vector of Hadamard measurement matrix to improve the efficiency of the algorithm and the result ofthe speech enhancement. Both objective and subjective experiments were employed tocompare the proposed approach with the sub-space method and the spectral subtractionmethod. Experimental results showed that proposed method outperforms other methods withthe highest PESQ, ABX and MOS score for Gaussian white noise and most kinds of colornoise.