Speech-Stream Detection Based On The Hilbert-Huang Transform

The analysis of non-stationary signal is one of the important contents in the fields of signal processing. But at present most of the methods used for processing the non-stationary signal is based on the theory of the Fourier analysis, so they also suffer the same shortcomings. Due to the non-linear characteristics of speech signals, the performance of traditional speech processing techniques based on linear methods cannot be improved any more. The lately developing and ameliorating non-linear, non-stationary signal processing theories have brought new direction for speech signal processing.The Hilbert-Huang transform (HHT) can cast the fetter of Fourier analysis and demonstrates the unique advantage in practical application. This new time-frequency analysis method has taken as a scientific breakthrough to Fourier analysis. Based on the speech-stream detection and the research of Hilbert-Huang transform under adverse background noises, the main study contents are as follows:The empirical mode decomposition method and Hilbert-Huang transform algorithm is elaborated detailedly. The main theoretical basis, the basic concepts and the algorithm are introduced. Besides, the properties of Hilbert-Huang transform are analyzed to give a deep insight of the transformation itself. By the comparison between Hilbert-Huang transform and Fourier transform, the advantage of Hilbert-Huang transform in processing the non-stationary signals can be indicated.Aiming at the problem of end-effect in the EMD, the mechanism of end effect is elaborated detailedly. Because the end-effect problem belongs to the small sample study problem, a novel method based on support vector regression machines is proposed to extend the end data. For support vector regression machines, different parameters are a very difficulty to select. In order to solve this problem, particle swarm optimization is introduced to support vector regression machines. Experiments show support vector regression machines based on particle swarm optimization can solve the end effects of HHT effectively.An algorithm for speech-stream detection in noisy environments, based on the Empirical Mode Decomposition (EMD) and the statistical properties of higher-order cumulants of speech signals is presented. With the EMD, the noise signals can be decomposed into different numbers of IMFs. Then, the fourth-order cumulant (FOC) can be used to extract the desired feature of statistical properties for IMF components. Since the higher-order cumulants are blind for Gaussian signals, the proposed method is especially effective regarding the problem of speech-stream detection, where the speech signal is distorted by Gaussian noise. Besides that, with the self-adaptive decomposition by the EMD, the proposed method can also work well for non-Gaussian noise.The speech detection based on Hilbert spectrum is content researched in this paper. With the EMD, the noisy speech signals can be decomposed into different numbers of IMFs. The main components of the voice signal are extracted by the marginal spectrum of each IMF components, and then marginal spectrum of time is used to extract the feature to detect the speech signal. Experiments show that the algorithm is robust to different types' background noises and capable to improve the performance of speech-stream detection.