Research On Calibration Methods For Distributed Microphone Arrays

With the popularity of portable devices such as smartphones and laptops,it is becoming more and more easy to establish a distributed microphone network in daily work and life.Compared to traditional fixed-topology microphone arrays,distributed microphone arrays have flexible network architecture,wide spatial coverage,strong fault tolerance and low-cost distributed data processing capabilities.However,these advantages also bring about challenges for the algorithms based on the distributed microphone arrays.Algorithms for traditional microphone arrays are not suitable for the distributed microphone arrays,it is necessary to design dedicated signal processing algorithms for distributed microphone arrays;many signal processing methods for distributed microphone arrays rely on the spatial position of nodes,but the node positions are usually unknown;the intra-circuits of the microphones in the distributed microphone network are different,which result in the mismatches among the frequency responses of microphones;nodes in the network have individual processing units,but the asynchronous problem for the independent clock systems in these processing units must be faced with.In this dissertation,the corresponding calibration methods and solutions are proposed to deal with the mismatch problems of sampling rates,microphone frequency responses and node positions in the distributed microphone networks.These methods can solve these mismatch problems,successfully,and achieves good ability in abating the influence of these mismatch problems on the signal processing techniques of distributed microphone arrays.The main innovations of this dissertation are as follows:(1)In dealing with the sampling rate mismatch problem in distributed microphone networks,a sampling rate calibration method for distributed microphone networks is proposed based on the multiple signal classification(MUSIC)spectrum estimation.When the cosine signal is used as the sampling rate calibration signal,the Hadamard product between the received signals of two node in the distributed microphone network contains two frequency-divisible cosine signals,and the frequencies of these two cosine signals are related to the sampling rate difference between these two nodes.Then,these two cosine signals are extracted by low-pass and band-pass filters respectively,and the spectrum of the lower-frequency cosine signal is estimated by the MUSIC algorithm to obtain the absolute value of the sampling rate difference.By observing the spectrum of the cosine signal with the higher frequency,the sign of the sampling rate difference is acquired.The sampling rate difference between nodes is finally compensated by a fourth-order Lagrange interpolation algorithm.This method can achieve a high sampling rate difference estimation accuracy with a low computational complexity.Simulation and real-world experimental results show that the proposed method is robust to noise and reverberant environmental conditions.(2)In calibrating frequency response mismatches among a small number of microphones,a microphone frequency response calibration method based on adaptive filters is proposed.The signals received by the microphones are similar when these microphones are gathered,and the deviation between the frequency responses of two microphones corresponds to the deviation between the output signals of these two microphones in this case.According to this feature,the microphones are gathered before the distributed microphone network is built,and adaptive calibration filters are introduced in the output terminals of microphones.Then,the frequency response mismatch problem is modeled by minimizing the joint mean square error among the spectrums of outputs of the adaptive filters.Finally,the frequency domain least mean square algorithm is used to obtain the filters for calibrating the frequency responses of these microphone in the optimization stage.Compared to existing microphone frequency response calibration methods,this method requires neither a reference signal nor information about the relative geometry position between the sound source and microphones.Simulation and real-world experimental results show that the proposed method still has good microphone frequency response calibration performance even under severe reverberation and ambient noise conditions.When a distributed microphone network contains a large number of microphones,even if the microphones are gathered,the correlation among the received signals of the microphones is weakened due to the position deviation among these microphones,but the correlation between the received signals of neighboring microphones remains good.According to this property,a microphone frequency response calibration method is proposed based on the consensus strategy in this paper.The average consensus strategy in the distributed data fusion is introduced that the frequency responses of neighboring microphones are iteratively fused to solve the frequency response mismatch problem.This method can solve the frequency response mismatch problem in the large-scale microphone networks.Since the signal processing and related operations only occur among neighboring microphones,this method can be realized with distributed computing of each node.Simulation and real-world experimental results show that the method achieves good frequency response calibration performance when the number of microphones in a distributed microphone network is large.(3)For the problem of node position calibration in distributed microphone networks,a three-dimensional node geometry position calibration method in distributed microphone network based on the direction of arrivals(DOA)is proposed.The theoretical DOAs of the sound source relative to the nodes are derived according to the three-dimensional rotation matrices and the translation vectors,and the corresponding measured DOAs are estimated by the time delays of the signal arrive the microphones in nodes.The node geometry calibration problem is then modeled as an optimization problem that solves the minimum mean square error between the theoretical and measured DOAs.Finally,the artificial bee colony(ABC)algorithm is applied to solve this optimization problem,and the position and rotation angle of each node are obtained.Compared with existing methods,the proposed method does not require information transmissions among nodes,which is easy to implement distributed computing.Instead of the gradient-based algorithms,the ABC algorithm is used to solve the cost function,which effectively avoid the iteration getting trapped into the local minimum value in solving the optimization problem.Simulation and real-world experimental results show that the proposed method has good three-dimensional node position calibration performance in noise and reverberation environments.Sampling rate mismatch,frequency response mismatch and node position calibration problem are widely presenting in distributed microphone networks,which seriously influence the performances of signal processing methods based on the amplitude and phase informance of node output signals or node position information.Therefore,before speech enhancement,speaker identification and tracking with distributed microphone networks,the these problems have to be solved to avoid the performance degradations of subsequent signal processing algorithms caused by the mismatches in the distributed microphone networks.