Hilbert-Huang Transform And Its Research On Speech Enhancement

Speech signal is highly nonstationary signal in that its second-order statistics (powerspectrum) change over time. Its spectral characteristics are fairly stationary over a sufficientlyshort period of time. Speech signal is usually disturbed by noise， which decrease the qualityof speech. Hence， we need to use speech enhancement technology to suppress noise， keepmore Intelligibility and increase quality. Speech enhancement technology is applied inwireless communication and teleconference. In this article， we summarize classic speechenhancement algorithm， derivate them and experiment on computer. Otherwise， weintroduce a novel speech enhancement method based on HHT and two methods based onEMD. Experimental results prove their effectiveness and performance in low SNRenvironment.The main contents and innovative points mainly include the following respects:1. We present history and research status of speech enhancement， study basic theory andalgorithm of speech enhancement. Spectral subtraction， MMES， signal subspacemethod and wavelet method are detailedly derived， which through computer simulationare compared about effect of enhancement.2. Hilbert-Huang transform is studied. Through analysis of linear frequency modulatedsignal， we summarize HHT’s characteristic of time-frequency analysis. A new principalcomponent method of energy is presented and applied in estimation of LFM signal’sparameters， while the method is expanded into speech enhancement field， whichprovide another research interest.3. At last， we discuss empirical mode decomposition’s advantage and drawback，summarize recent EMD based speech enhancement methods， present two novel EMDbased algorithm: EMD-SS and EMD-Mband. They are derivative instruments of spectralsubtraction and multi-band spectral subtraction respectively. Computer simulation ofEMD-SS and EMD-Mband are carried out in MATLAB. Speech quality andIntelligibility are compared among spectral subtraction， multi-band spectral subtraction， EMD-SS and EMD-Mband. The results show that EMD-SS has better speech quality andEMD-Mband has better speech intelligibility in low SNR environment.