Single Channel Speech Enhancement Algorithm Based On Blind Source Separation

In the actual process of communication, voice is often disturbed by ambient noise, thereby affecting the transmission quality of voice signal. Reduction of the voice signal quality will further affect the performance of the subsequent signal processing system, or even result in more serious consequences. Therefore, the received noisy speech signal needs to be enhanced to effectively suppress ambient noise, improve the quality of voice signal. Blind source separation, as a hot research topic of modern signal processing has a solid theoretical foundation and broad application prospects in the field of speech enhancement. Due to the assumption that the number of observed signals is not less than that of independent source signals in blind source separation algorithm, the application of blind source separation in single-channel speech enhancement is studied on the basis of previous work in this paper, thus can restore the pure original speech as far as possible from the noisy speech.In order to solve the problem that the traditional blind source separation algorithm can not be directly applied to single channel speech enhancement, propose a method to obtain another way of virtual observed signal using singular spectrum analysis. Then use improved algorithm with combination of temporal predictability and wavelet transform for separation, effectively improving the signal to noise ratio and auditory perception of the enhanced voice signal.The constructed virtual observed signal on the basis of the original noisy signal will affect the independence between the two observed signals. So provide the two-dimensional input of ICA transform by Bark-scale wavelet packet decomposition of noisy speech, then select the appropriate threshold function, and improve the old algorithm using the threshold method, greatly reducing the complexity of the algorithm, and achieving good effect of enhanced voice.In addition, this paper studies the principle of single channel speech enhancement based on noise estimation and blind source separation, then propose a dynamic estimation algorithm of noise based on the update of short-term noise spectrum. According to the proportion of noise in the noisy speech signal to dynamically adjust the estimated coefficient of the short-term noise spectrum, in order to better estimate the noise in the noisy speech. Take the estimated noise signal as another way observed signal input of the improved algorithm based on information maximum, to achieve the blind separation of pure voice and noise signals. Experiments show that the proposed algorithm can effectively eliminate the white Gaussian noise and colored noise, and obtain enhanced speech signal.