The Application Of Hilbert - Huang Transform In Noisy Speech Processing

Speech is the most important communication tool, human society and human thinking tool, the human society civilization is closely related to itSpeech language information, people’s information interaction mainly by voice.In the actual speech environment, more or less there are varying degrees of environmental noise interference.Noise makes the voice quality to drop, voice signal processing device can not work normally, information exchange, with noise speech denoising to become a problem to be solved.Traditional various time-frequency analysis techniques such as Fourier analysis is a stationary signal oriented, not suitable for analysis of non-stationary signal, the non-stationary signal is the typical representative of speech signal, it is a time-varying random signal.Developed in recent years it is this article main research to on-stationary signal Hilbert-huang transform theory to process the voice signals with noise, noise, so as to achieve the speech enhancement.Voice signals with noise by the empirical mode decomposition, the gens IMFs intrinsic mode components.The IMF points that make up the IMFs frame to calculate their so-called smooth variance.Based on the variance by the initial of IMS itself, combined with the IMFs global decision whether speech frame, and on the basis of the further detect the speech endpoint.This is the first step in speech enhancement, determine the target object, remove unnecessary of speech processing.For the detection of speech signals with noise after empirical mode decomposition, the noise distribution and voice in the IMF, we give different rights to be enhanced or suppressed. Still contain different level after further to deal with the noise of each IMF component, using adaptive threshold adjustment to further suppress the noise, the reconstructed speech finally.In this paper, the speech made by additive noise pollution.The enhancement and comparative experiments show that, the author proposed method has better performance on all kinds of noise, and wavelet analysis is only suitable for the white noise and poor performance under low SNR.