Application Of HHT In Speech Eenhancement And Speech Endpoint Detection

With the development needs of machine translation, man-machine dialogue anddialogue between machine and machine, speech signal processing technology hasbecome research focus in digital domain.Speech signal is unavoidably disturbed by the noise in surrounding environment.How to reduce the influence of noise to improve output effect of speech signalprocessing system becomes very important. Speech enhancement is a specializedtechnology which is used for extracting the pure voice from the noisy speech signal asfar as possible, thereby reducing the impact of noise. When the speech signalprocessing is in the ideal state, voice segment is wanted signal and non-voice segmentis unwanted signal. But in noisy environment, voice is covered up by noise, so onlyextracting accurate voice segment through speech endpoint detection can guaranteethe feasibility of subsequent signal progressing. Therefore, speech enhancement andspeech endpoint detection are important parts of speech signal processing technology.Speech signal is a kind of complex non-liner and non-stationary signal. And,Hilbert-Huang transformation (HHT) is a new algorithm which is suitable forprocessing non-liner and non-stationary signal. The main works carried out in thispaper are speech endpoint detection and speech enhancement in strong noisyenvironment, and the concrete contents are as follows:1) Through analyzing the causes and impact of boundary effect of empiricalmode decomposition (EMD) and considering the distribution of extreme point ofsignal, a processing method of boundary effect based on the best matching method ofgroup of extreme point is proposed. Simulation experiments prove that boundaryeffect can be efficiently restrained using the proposed approach.2) Speech endpoint detection based on HHT is studied in this paper. In order tousing the multi scale filtering characteristics of EMD and the distributioncharacteristics of intrinsic mode function (IMF), an approach of speech endpointsdetection based on the improved HHT in strong noisy environment is proposed. Themain steps of this method are as follows: firstly, the signal to noise ratio (SNR) ofinput speech of endpoint detection system is raised by twice noise reduction. Then theselected IMFs are weighted to make the characteristic of voice segment clear. The results of simulation experiments indicate that the proposed approach not only canimprove correctness of speech endpoint detection, but also can improve its robustness.3) In order to improve the effect of speech enhancement based on EMD andeliminate “music noise” produced by speech enhancement based on self-adaptivefilter, an approach of speech enhancement combining EMD and self-adaptive filter isproposed. This approach combines multi scale filtering characteristics of EMD withthe strong usability and simple realization of self-adaptive filter to improve the effectof speech enhancement. And the proposed method can solve the above problemsefficiently by subjective audition and observing picture in time domain, spectrogramand curve of the increase of SNR.