Research Of Voice Activity Detection And Speech Enhancement Based On Empirical Mode Decomposition

Voice activity detection and speech enhancement are front speech signal processingmethods, and their accuracy will determine the subsequent processing such as speechrecognition, speech coding results largely. Effective voice activity detection method can notonly reduce the amount of data storage, shorten the signal processing time, but also caneliminate silent clips of the noise; speech enhancement can improve the quality of voicesignal, improve signal intelligibility and help listener to understand. Now, the commonly usedmethod can achieve good results in weak background noise, but the noise components aremuch more in realistic environment. Whether can detect the endpoint and enhance the speechsignal effectively have the important meaning in low Signal to Noise Ratio (SNR) conditions.As a new signal processing method, Empirical Mode Decomposition (EMD) shows thegreat superiority in non-stationary signal analysis, it has become an effective tool fornon-stationary signal processing. Speech signal is a typical non-linear and non-stationarysignal, the time-frequency analysis based on EMD provides a new means and methods forspeech signal processing, and open up a new effective way for voice activity detection andspeech enhancement in noisy environment. This paper start in-depth research and proposenew effective method for speech signal detection and enhancement base on EMD, the mainwork is as follows:1. This paper describe the basic knowledge of the speech signal processing, analyse theexisting voice activite detection and speech enhancement method; in allusion to EMDalgorithm make a further research, discusses the basic characteristics and existing problems,use EMD algorithm to analyse the fine time-frequency structure of speech signal.2. The Hilbert energy spectrum curve of speech signal is fluctuate in strong noiseenvironment，it has a great influence to voice activity detection. So an effective voice activitydetection algorithm is proposed based on Hilbert-Huang Transform (HHT) and OrderStatistics Filter (OSF) in this paper. This method first decompose noise signal into intrinsicmode functions by EMD. Then the Hilbert energy spectrum is synthesized by adaptive weightselection of each intrinsic mode functions, through OSF to smooth the energy spectrum.Finally, the speech and noise divergence is judged by means of the smoothed energy spectrum. Experimental results show obviously that under complex noisy environment, this method isstill able to effectively detect the speech signal, and reduce the error detection rate in lowSNR conditions.3. Due to the decrease of EMD quality and the distort of voice harmonic after denoise inlow signal to noise ratio cases, an speech enhancement method was proposed based on EMDto improve the speech intelligibility. This method firstly through the adaptive noisecancellation technology to pretreatment the speech signal, eliminated interference of whitenoise and promote EMD quality, then got different scales of intrinsic mode functions by EMD,chose different denoising method to high and low frequency intrinsic mode functions, finally,used the harmonic regeneration method to reconstruct the enhance speech signal, perfectedthe lost high times harmonic composition. Experimental results show obviously that thismethod can not only improve the SNR but also promote speech signal quality, furtherstrengthen the speech signal intelligibility.