Indoor Speech Separation And Sound Source Localization Based On Dual Microphone

With the continuous development of signal processing technology for microphone arrays,various new theories and methods have been proposed.Nowadays,microphone array technology has been widely used in teleconferencing,indoor and outdoor voice communication,human-machine voice interaction and other systems.The dual-microphone-based indoor multi-source information sensing system has the characteristics of small size,low power consumption,low cost,and the like,and is more suitable for the development trend of miniaturization of intelligent products.However,the system based on dual microphone collects fewer signal samples than the number of sound sources(underdetermined problems),relatively little spatial information,and relatively few other information available.How to fuse limited information in underdetermined situations,become the focus of research.Therefore,this paper discusses the speech separation and sound source localization of underdetermined mixed signals using dual microphone.Based on the time-frequency characteristics of speech signals and the spatial position information of dual microphones array,an indoor speech separation and sound source localization is proposed.system.The main work of this paper is:1.In order to further improve the separation quality of source signals,the traditional DUET(Degenerate Unmixing Estimation Technology)algorithm is improved.Firstly,for the problem of inaccurate estimation of the mixed parameters,a multi-resolution Common Fate Transform(MCFT)transform is used instead of the STFT(Short-Time Fourier Transform)transform to construct a two-dimensional histogram.Secondly,for the problem that part of the time-frequency point is lost when the ideal binary time-frequency mask is separated,the Gammatone filter is used to smooth the binary time-frequency mask to better reconstruct the source.In this thesis,SDR(Source to Distortion Ratio),SAR(Source to Artifacts Ratio)and SIR(Source to Interferences Ratio)are used as evaluation metrics.The experimental results show that the improved DUET algorithm has significantly improved separation performance compared with the existing algorithms.2.In order to explore the possibility of using dual microphone for multiple sound sources separation and positioning in the two-dimensional plane,an indoor voice separation and sound source localization system is proposed.Based on the signal collected by the microphones,a two-microphone time delay-attenuation model is established.Then the DUET(Degenerate Unmixing Estimation Technique)algorithm is used to estimate the model's delay-attenuation parameters,and a parameter histogram is drawn.In the speech separation stage of the system,BTFM(Binary Time-Frequency Masking)is established.According to the parameter histogram,binary speech masking is used to separate the mixed speech.In the sound source localization stage,the mathematical equations that determine the position of the sound source are obtained by deducing the relationship between the model attenuation parameters and the signal energy ratio.The Roomsimove toolbox is used to simulate the acoustic environment of the room.Through MATLAB simulation and geometric coordinate calculation,this paper completes positioning in the two-dimensional plane simultaneously with separation of multiple sound sources.Experimental results show that the positioning error of the system for multiple sound sources is below 2%.