Modeling And Localization Based On Microphone Arrays

Microphone arrays, which are commonly employed to capture acoustic signals, are an important part of various acoustic communication systems. Microphone arrays are extensively applied to speech enhancement, beamforming, system identification, dereverberation, speech recognition and speaker recognition, sound source localization and tracking, acoustic echo cancellation, and speech separation. This dissertation focuses on research on blind multichannel identification, source localization and tracking based on microphone arrays.Blind multichannel identification (BMCI) is to estimate the channel impulse responses of an unknown multichannel system based only on the output signals. To improve the resilience of BMCI to non-Gaussian noise, we propose a robust normalized multichannel frequency-domain least-mean M-estimate algorithm. Unlike the traditional approaches that use the squared error as the cost function, we use an M-estimator to form the cost function, which is shown robust to non-Gaussian noise with a symmetric α-stable distribution. Simulations demonstrate the superiority of the proposed algorithm.To localize sound sources in room acoustic environments, time differences of arrival (TDOAs) between two or more microphone signals must be determined. In this dissertation, we partition the joint entropy of multiple random variables into two classes of information: mutual information shared by the multiple random variables and non-mutual information among them. We extract the non-mutual information among an array of microphones to estimate TDOA. Simulations in reverberant environments justify the effectiveness of the proposed algorithm.The multichannel cross-correlation-coefficient (MCCC) algorithm, which is an extension of the traditional cross-correlation method from two-to multiple-channel cases, exploits spatial information among multiple microphones to improve the robustness of time delay estimation. In this dissertation, we propose a multichannel spatio-temporal prediction (MCSTP) algorithm, which can be viewed as a generalization of the MCCC principle from using only spatial information to using both spatial and temporal information. We also propose a recursive version of this new algorithm, which can achieve similar performance as MCSTP, but is computationally more efficient. Experimental results in reverberant and noisy environments demonstrate the advantages of this new method.Finally, we use directional microphones to construct a square array, and analyze the frequency response and directionality of this array. To make the array capture the sound from all directions, we analyze how to design this array. Based on the square array, we propose a source tracking algorithm with the ability to track the speaker with the maximum speech power. Experimental results in anechoic and general rooms demonstrate the effectiveness of the algorithm.