Research On Hilbert-Huang Transform And Its Application In Speech Enhancement

Speech enhancement is a method of extracting speech information from noisy speech signal, which is extensively used in speech recognition, speech coding, speech communication and other fields. The approaches of speech enhancement are numerous. In the past we used to enhance the speech signal by the traditional methods, including time-domain, frequency-domain, windowed Fourier transform and wavelet transform, but there are shortcomings that limit their applications. Hilbert-Huang transform is a new and powerful theory for the nonlinear and non-stationary signal analysis, which has important theoretical value and wide application prospects. It has effectively been applied in digital signal denoising domain and plays a more important role.The commonly used time-frequency analysis methods for non-stationary signal are introduced firstly, and then the basic knowledge of Hilbert-Huang transform is summarized, as well as the studying background, existing problems, studying progress and its applications in various research areas particularly in the field of signal denoising. On the basis of that, Information entropy methods and principal component analysis are introduced to improve the method of Hilbert-Huang transform. Due to the disadvantages of stopping criteria for traditional empirical mode decomposition algorithm, an algorithm for Hilbert time-frequency spectrum entropy is given. Based on this, the new stopping criteria for the sifting and decomposing is presented, and the algorithm for empirical mode decomposition is improved as well. The improved algorithm for empirical mode decomposition could get more accurate results and overcome the problems of false mode and modal mixture in Hilbert-Huang transform to a certain extent.The enhancement theories as well as methods of speech signal which is polluted by additive noise are studied, and the decomposing characters of both speech signals and noise signals are analyzed. A speech endpoint detection algorithm and a new speech enhancement method of adaptive multi-scale and multi-threshold based on Hilbert-Huang transform are presented. Using the EMD algorithm, firstly the speech signal is decomposed into several IMFs, then speech detection and denoising is done selectively according to their own characters, and lastly the signal is rebuilt. While the SNR of the speech is low, the experimental results show that the denoising effect of the proposed method is better than that of other methods based on wavelet shrinkage, the algorithm is valid on several noise conditions for most of speech signals and is capable to improve the SNR of the speech.