Feature Extraction And Recognition For Sound Source Localization Using A Small-Sized Microphone Array

Sound source localization using microphone array is an important technology in the field of speech signal processing and widely used in practice.In some practical application,such as in round-table video conference,all possible sound source locations are only confined to some discrete areas.Therefore,it is reasonable to deal with the sound source localization problem from a machine learning point of view.Machine learning-based sound source localization methods have considered the prior information of source locations and the key to these methods is how to extract source location features.The existing feature extraction methods,for example,the time difference of arrival features between microphones,cannot meet the requirement of localization accuracy when using a small-sized microphone array in reverberant and noisy environments.This dissertation studies the feature extraction method of sound source localization using small-sized microphone array in a room by combining the analysis of first-order harmonic and machine learning theory.In addition,recognition methods also have effects on the performance of the localization method.By the deep learning technique of denoising autoencoder,this dissertation studies the sound intensity feature-based recognition method for sound source localization using a small-sized microphone array.The main works and contributions of this dissertation are summarized as follows:Firstly,the typical feature extraction method based on the time-difference-of-arrival are studied.Simulation experimental results show that this method can realize sound source localization under circumstance of larger array size,lower reverberation and higher signal-to-noise ratio(SNR).However,when we use a small-sized microphone array,the localization accuracy of this method decreases in some degree and its localization performance is unstable.Secondly,a robust source location feature extraction method based on sound intensity is proposed.This method is on the basis of sound intensity estimation and includes three steps.The first step is to calculate every time-frequency instantaneous sound intensity with phase transform technique to improve its performance of anti-reverberation.The second step is normalizing each sound intensity component to avoid the mismatch between the feature vectors which caused by the diversity of human speech.The last step is extracting the sound intensity of subarrays as redundancy feature information to increase the error-tolerant rate and then improve the localization accuracy.Experimental results show that the proposed feature extraction method has higher localization accuracy and it is robust in reverberant and noisy environments,and also suitable for small-sized microphone arrays.Thirdly,a deep learning based recognition method which is suitable for sound source localization using a small-sized microphone array is proposed.This method uses the sound intensity information as features,and the denoising autoencoder is primarily given unsupervised training with all possible sound source location feature data.Then,supervised method is used by the back propagation algorithm to tune parameters of the network.Finally,feature vectors of newly-acquired data are put into the well-training network in order to achieve the sound source locations.Compared to the existing recognition method,denoising autoencoder has stronger anti-interference ability.Experimental results show that the proposed method has higher recognition rate in the condition of long reverberation time and low SNR and definitely has advantages in small-sized microphone array sound source localization.