Convolution Model Based Sound Source Separation Algorithm Research Under High Reverberation

Nowadays,sound source separation has become a research hotspot in the field of speech signal processing,which is widely used in smart homes,teleconferences,and noise-canceling headphones.Blind source separation(BSS)is often the case for sound source separation,and the so-called blind means that using only information about their mixtures observed in each input channel to estimate source signals.The estimation is performed without possessing information on each source,such as its frequency characteristics and location,or on how the sources are mixed.Many theories and methods have been proposed to solve the problem of over-determined or determined BSS,while the problem of BSS under under-determined conditions is a hotspot and difficulty of current research.This thesis studies the problem of under-determined BSS in a reverberant environment,and proposes a BSS algorithm for convolutional mixed signals under under-determined conditions.Then this thesis improves the source separation method based on deep learning,and apply it to the single-channel sound source separation problem.The specific work is listed as follows.1.Reverberation and noise often exist in the real world,which is also the direct cause of the sharp decline in the performance of most BSS algorithms.In order to improve the robustness of the algorithm against reverberation and noise,this thesis proposes a convolution model-based under-determined BSS algorithm under high reverberation.In this algorithm,the mixing system and the source signal are jointly estimated.First,starting from the instantaneous narrowband model in the time-frequency(TF)domain,a regularized optimization framework is designed.The final optimization problem is solved in an alternate optimization manner,using the alternating direction method of multipliers(ADMM)to find the corresponding solution.Then the model is extended to the convolutional narrowband model in the TF domain to handle high reverberation scenes.The experimental results show that compared with the latest methods based on signal processing,the algorithm proposed in this thesis has obvious performance advantages.2.The problem of sound source separation under single channel condition is more challenging,because only one channel of information is available and there are no constraints,producing countless solutions.This thesis introduces deep recurrent neural network(DRNN)as a regression model to find the deep structure and regularity of signal reconstruction from a mixed signal containing two sources.In order to alleviate the problem of interference between different sources,this thesis integrates the scaleinvariant signal-to-distortion ratio(SI-SDR)loss function and discriminative network training criteria to form a new training objective function.The experimental results show that this method demonstrates a better separation performance compared with other methods.