Speech Denoising Methods Based On Tensor Model

Speech is one of the most common and convenient ways of communication in nature,and it is also an important means of information transmission.However,in real environments,the existence of background noise and interference not only reduce the intelligibility of the speech,but sometimes even loses part of the speech information.It also brings a series of challenges to subsequent speech processing system,such as conference transfer and recording system,speech coding and automatic speech recognition systems.Therefore,in the noisy real environments,the research of speech enhancement is extremely important,which is also an important pre-processing part in speech signal processing system.Speech enhancement means that when the speech signal is interfered by background noise or interference,it can suppress or attenuate noise and interference to the greatest extent possible while ensuring that the speech signal is as complete as possible.In this way,we can obtain as far as the pure speech signal from the received mixed noisy speech signal.Hence,the quality and intelligibility of these contaminated speech signals can be improved.According to the number of microphones,the speech enhancement algorithms can be divided into two categories:single channel and multi-channel speech enhancement.The traditional single-channel speech enhancement system is widely used in engineering project because it is relatively easy to implement and has a certain noise reduction effect.However,with quite loud noise or interference,single-channel speech enhancement system may introduce"music noise"or cause speech distortion.For the multi-channel speech enhancement algorithms,aside from getting the time-frequency domain information of the speech signal received by the microphone array,the spatial information of the multi-microphone signal can be also utilized,thereby better noise reduction effect can be achieved.Most of these existing multi-microphone speech denoising algorithms directly utilize the correlation of frequency-space domain or time-space domain of the received noisy speech signal.However,little attention has been paid to joint exploitation of correlations in the time,space and frequency domain.In this paper,we propose to integrate joint time-space-frequency information of the observed speech signal to represent the received multi-microphone speech signal as a three-dimensional tensor.Then,the multi-mode filters are established by the tensor analysis tool like alternating least squares method.The three filters of time,frequency and space domain are used to filter the received noisy speech signals iteratively,as a result,the noise is suppressed to recover the desired clean speech signal.In order to improve the performance of the multi-dimensional filtering algorithm,we will further combine this method with the traditional beamforming algorithm.We first perform noise reduction preprocessing on the input noisy signal under the high-order tensor frame,and eliminate part of the background noise.Secondly,the obtained pretreatment of multi-channel speech signal is then processed by beamforming algorithm to further eliminate directional interference.The combination of these two methods not only has a good suppression effect on the non-directional noise such as self-noise of the microphone,but also can well suppress the directional interference.That is,the proposed frame is more applicable to the actual environment.Experiments are conducted to test performances of the proposed framework on both the simulated and realistic acoustic systems.The experimental results of the simulation environment system and the real acoustic system show that the proposed framework has an advantage over other tested algorithms in terms of subjective and objective measures.