Research On Binaural Sound Source Localization Based On Convolutional Neural Network

Most sound source localization(SSL)algorithms are based on multi-microphone arrays.However binaural sound source localization has advantages of compactness and low-cost.Traditional binaural SSL algorithms are susceptible and perform poorly in reverberation and noise environments.Convolutional Neural Network(CNN)modeling has high level of robustness,significantly improved the performance in complex environments.In order to cope with the challenges of complex environments,this thesis proposes two binaural SSL algorithms based on CNN: binaural SSL algorithm based on Densely Connected Convolutional Networks(DCCN)classifiers,and binaural SSL algorithm based on Soft-argmax regressor.(1)The binaural SSL algorithm based on DCCN classifier is developed from the traditional CNN models.This algorithm introduces a dense connection structure.The character of feature reuse makes the model have powerful fitting ability without huge parameter amount.And the 1 × 1 convolution in DCCN also effectively compresses the model size.Speech with noise and simulated reverberation is used for training.Speech with noise measured reverberation is used for testing.Through such experimental settings,the algorithm’s performance in unseen environment is verified.Experiments show that the localization accuracy of this algorithm is significantly better than traditional algorithms in complex environments.(2)The binaural SSL algorithm based on Soft-argmax regressor is developed from the neural network classifier.This algorithm introduces Soft-argmax regressor structure.The algorithm can transform any neural network classifier with Softmax into a regressor,which will improve the performance in bad environments.In experiments section,we transform the DCCN classifier into a Soft-argmax regressor.Experiments show that the root mean squared error(RMSE)of this algorithm is less than that of the classifier model and traditional algorithms in complex environments.