Acoustic Source Localization Research Based On Machine Learning

Acoustic source localization method based on microphone array signal processing technology has been widely used in many fields,such as video conference,speech enhancement,intelligent robot,intelligent home,etc.Due to various kinds of interference,however,the localization performance will be degraded,even those methods can't work,especially the reverberation,noise and other adverse factors often exist in indoor environment.For acoustic source localization,therefore,how to improve the the robust ability and the localization accuracy in harsh conditions is a research focus.In recent years,a classification recognition method based on machine learning algorithm is used to locate the sound source and which has been gotten attention.This kind of method not only has stronger robustness,but also can still work when the microphone can't receive direct sound compared with the traditional sound source localization algorithm.How to enhance the performance of the source localization in a reverberation and noise environment is researched based on machine learning algorithm in this thesis.First of all,the acoustc propagation model and the microphone array signal model are analyzed,and the traditional sound source localization algorithms GCC and SRP-PHAT as well as the machine learning algorithms were introduced.On this basis,using LDA classifier to identify the GCC-PHAT function is put forward in this thesis firstly.And the simulation results show that the locating performance in serious reverberation condition is better than that of Naive Bayes classifier.Secondly,the feature of cross-correlation function was transformed by LDA.The simulation results show that the locating accuracy using the projection transformation by LDA under the harsh environment is much higher than that of before transformation.Thirdly,single classifier research will be generalized to multiple classifiers.Adaboost and Bagging methods were used to ensemble multiple classifiers which show higher locating accuracy than single classifier.Finally,the Bagging method will be optimized using the k-means method to selective ensemble the individual classifier which enhances the robust ability of acoustic source localization further.