The Research Of Speaker Localization Method Based On Microphone Array Under The Strong Noise Enviroment

Speaker localization and tracking based on microphone array is an important research topic in human-computer interaction, it has been widely applied in the field of robotics and intelligent conferencing systems. Consider the problem of the low accuracy and poor robustness for single-feature acoustic localization based on standard particle filter, the paper takes full advantage of the complementary between different characteristics of audios, and improves a new method for speaker localization based on adaptive multiple features fusion and improved particle filter. The method can be used to achieve more robust and accurate speaker location under the strong noise environments. The main work and innovation of this paper is as follows:Taking into account the low accuracy of the standard particle filter which lost the measurement because of taking the prior probability density as the proposal distribution function, the study proposed a new distribution function using Iterative Kalman to update the mean and variance produced by Unscented Particle Filter, integrated the latest audio measurement information to the proposal distribution. The new proposal distribution enhanced the real-time correction of the latest information to the system model and improved the transfer accuracy of the system model. Simultaneously the improved particle filter was used for speaker localization and presented an acoustic source localization method based on Iterative Unscented Particle Filter. Simulation results show that the improved method is more accurate compared to the particle filter and unscented particle filter.Acoustic source localization based on single feature can be easily affected by background noise, and thus leads to low positioning accuracy and robustness. In order to solve this problem the fusion of feature information was introduced and proposed a speaker localization method based on multi-feature fusion. Firstly, the method constructed a speaker localization system model, and then used it to define difference function between characteristics. The function can evaluate the targeting support consistency of different characteristics. Secondly, a new adaptive feature fusion strategy was been proposed by analyzing advantages and disadvantage of plus fusion and multiplicative fusion. The difference function will be compared to a defined threshold and decided which fusion strategy can be used. Finally, the above adaptive feature fusion strategy was used to fuse two characteristics of steered beamforming and phase transform steered response power. By applying it to a speaker localization system based on improved particle filter, the performance of the new method was compared with single-feature based localization system under different trajectories. Simulation results show that the fusion results had better use of complementary information between different features and improved the accuracy of the speaker localization system.