Study On The Methods Of Time Delay Estimation And Speech Enhancement

Speech signal processing is one of the key techniques applied in many new fields such as information highway, multimedia technology, office automation, modem communication technology and artificial intelligence systems. But speech signals are inevitably disturbed by various interferences. As a result, the signals received by microphone are the speech corrupted by noises instead of the pure speech. As a useful means to restrain these interferences, speech enhancement is the basis of speech signal application.The main works in this paper are as follows:The traditional methods of Time Delay Estimation (TDE) are carefully studied. TDE is a key part in many microphone array based speech enhancement systems. However, some of them are just suitable for uncorrelated noises. Others are only suitable for correlated noises.d Besides, their performance degrades dramatically in the case of low Signal-to-Noise Ratio (SNR) and strong reverberation. Two improved algorithms are proposed in this paper. The one based on the theory of maximum a posterior estimation can improve the performance in the cases of low Signal-to-Noise Ratio. The other based on adaptive eigenvalue decomposition in Digital Wavelet Transformation (DWT) domain can yield good performance in various reverberation and SNR conditions. Computer simulation results show that the proposed methods outperform the original methods.When the number of sources is greater than the number of mixtures, the performance of the traditional Blind Source Separation (BSS) algorithm is poor. To solve this problem, a improved method combining DWT and BSS is proposed. Moreover, the computer simulation results indicated that the improved method yield better SNR improvement and better quality of the enhancement speech signal than the original method.Subband blind source separation and post-processing based speech enhancement method is studied. The performance of Adaptive Noise Canceller (ANC) which usually brings in distortion largely depends on the performance of BSS. And it is a complex problem to judge of the BSS performance. To solve these problems, some improvement has been made on it. The spectral subtraction instead of ANC is used as post-processing module to archive further enhancement. Simulation results show that the improved method dose less harm to the speechs, greatly improves the quality of speech and successfully depresses noise.