Recognition And Tracking Of Multiple Speakers Using Audio Features In Intelligent Conference Room

Using microphone arrays to pick up the speech signals,then extract the useful information that can reflect location of the speakers and keep tracking of the current position of the speaker is an important research field in signal processing and target tracking.This technique has a wide application in smart conferencing systems,speech enhancement,robot navigation etc.In real intelligent environments there are often more than one speaker,and the speakers’ locations and their numbers are constantly change.Therefore,the research of tracking and locating time-varying number of speakers has more application value.It can help to build intelligent conferences that are more practical and promote the interaction between human and computer.The main work and innovations of this thesis are as follows:(1)For the problem that the resampling part of the particle filter algorithm is difficult to be parallelled,the resampling results in low efficiency in the parallel execution of the particle filter algorithm.Therefore,the parallel strategy is introduced in the particle filter algorithm,and on the basis of the parallel particle filter algorithm,combining parallel prefix sum method to overcome the inter particle dependency of particle filter algorithm in resampling so as to solve the problem that the particle filter algorithm is difficult to be parallelled in the parallel implementation process.(2)Aiming at the problem of parallel computing complexity in parallel particle filter algorithm,this paper analyzes the memory access mode of parallel protocol algorithm in GPU,and there is a serious memory access conflict in parallel algorithm.In this paper,we use the parallel prefix method of filling address to add a fill in the index of each shared memory array,and achieving access to the improved shared memory,to solve the parallel protocol memory access in the existence of a serious memory access conflict,improve GPU hardware resource utilization and algorithm real-time.(3)For the measurements of particle filter,need to be associated with the respective target,and each target requires a separate particle filter in multiple acoustic source tracking systems,which will lead to the increase of particles and calculation of the system.This thesis uses the parallel particle filter combined with K-means clustering algorithm based on GPU to track the multiple audio targets.The particle filter predicts and updates the multi-target states as a whole.The K-means clusteringalgorithm based on GPU is used to cluster particles after resampling and classifies particles of the same target.The circular microphone arrays are also used to obtain positioning features because of its high ability to discriminate between the sound sources.Simulation results show that the above method effectively solves the data association problem of particle filter and improves tracking accuracy of multi-target.