The Time-frequency Analysis Of Speech Signal Based On Hilbert-Huang Transform

With the development of electronic computers and artificial intelligence, speech signal processing has become a hot research area nowdays. The speech signal is a kind of complex non-linear and non-stationary signal. Time-frequency analysis is an effective tool to analyze and process non-stationary signals, which shows a good distribution picture in the joint time-frequency domain. Using time-frequency energy distribution we can extract information in given time or frequency. Time-frequency analysis has brought new direction for speech signal processing. But speech signals are often corrupted acoustically by ambient noise. So speech enhancement and speech detection are strongly desired to improve the quality of the system.The traditional methods such as STFT and WVD exist window effects and cross terms respectively and are not adaptive. This paper studies a new signal processing method named Hilbert-Huang Transform which can effectively response the non-linearity and non-stationary signal and is applied on the speech signal processing. The main research works and contributions are as follows:1. The theory of HHT and the problems of HHT are studied, the appropriate method delt with end effect is adopted and is applied on the time-frequency analysis of speech signal based on HHT. Simmulation results which are compared with STFT, WVD and Choi-Williams distribution demonstrate it can gain more fine time-frequency structure.2. It can distinguish differences of time-frequency distribution of different tones exactly by Hilbert spectrum and marginal spectrum. We establish M-diatance according to the energy of Hilbert marginal spectrum under different tones. The preliminary test shows the effectiveness of this method.3. The alogrithm of speech ehancement based on Hilbert-Huangtransform and human auditory masking is developed in this paper, the high-frequency IMFs is processed with the human auditory masking. .Simulation results show that it can reduce the measure value of speech distortion and improve the SNR, speech articulation and intelligibility.4. The theory of TEO and double threshold method are studied, TEO reflects energy more reasonably and be more adaptive to FM or Am signals. We develop speech endpoint detection based on EMD and improved double threshold method, simulation results show this algorithm has better detection capability in low signal to noise ratio environments.