Research On Sound Source Posture Estimation Method Based On Machine Learning

With the development and promotion of human-computer interaction technology of mobile smart devices,the sound source localization techniques are widely used in the field of speech enhancement,video conferencing.However,the interference of background noise,reverberation and other factors in the indoor environment will lead to the decline of the system positioning effect.How to improve the robustness and accuracy of the sound source localization algorithm has been a hot topic of current research.At the same time,in the actual application process of smart home and other devices based on the sound source localization algorithm,the orientation information of the sound source in a specific area is often important.How to effectively use the characteristic information of the sound to improve the accuracy of the sound source orientation discrimination is also a difficult problem in current research.To solve the above problems,this paper deals with the pose estimation of the sound source from the perspective of machine learning,according to the position and orientation of the sound source and the following two parts have been completed:(1).For the sound source localization algorithm based on location fingerprints after preprocessing its database through K-means clustering and other algorithms in its preprocessing stage,the test points located at the edge of the cluster cause large positioning errors due to insufficient reference information.This paper proposes an optimization algorithm for local clustering to reduce the error of the test points located at the edge of the cluster and improve the positioning accuracy of the system.First,this paper uses principal component analysis(PCA)algorithm to reduce the dimensionality of time difference of arrival(TDOA)feature information collected at test points.Then,a local clustering optimization algorithm was used to select the clusters corresponding to the test points twice.Finally,the enhanced WKNN algorithm was used to estimate the sound source position of the test points.The experimental results show that the sound localization method based on the local clustering optimization algorithm can effectively improve the positioning accuracy of the test points located at the edge of the cluster and also improves the real-time and robustness of the algorithm.(2).Aiming at the existing sound source orientation estimation method when judging the direction of the sound source by manually setting the threshold value,there are great subjective problems due to the difference in operator level and the difference in threshold selection.This paper proposes a sound source orientation estimation method based on attention mechanism by using the long and short-term neural networks(LSTM)and attention mechanism to analyze the energy characteristic information of sound and achieve the direction of the sound source in a specific indoor area.The results of comparative experiments also show that using the attention mechanism can effectively improve the accuracy and generalization of the model and can also reduce the time overhead of model training.