Sound Source Localization And Recognition From Complex Background

With the development of new technologies such as smart speakers,voice assistants,smart offices,and robots.The demand for intelligent signal processing,especially sound signal processing,is increasing.The application of artificial intelligence in the field of signal processing is also becoming more and more widespread.Like machine vision,machine hearing will become an important application in the field of intelligent perception in the future.Therefore,in a complex environment,especially in an acoustic environment with ambient noise and indoor reverberation,the localization and identification of sound sources is of great significance for intelligent speech processing.The microphone array receives sound signals through multiple directions and plays a key role in estimating the direction of arrival(DOA)of the sound source and identifying and classifying overlapping sounds.Using a microphone array,this paper has studied and explored the location and identification of multiple sound sources in a complex acoustic environment.The main work and innovation are as follows:First,the mathematical model of the propagation and reception signals of the microphone array is constructed,and the GCC-PHAT and SRP-PHAT algorithms based on the time of arrival difference,and the MUSIC algorithm and ISSM algorithm based on spectral estimation are analyzed.And LCMV adaptive beamforming algorithms,simulations have achieved the performance of various algorithms.Secondly,the experiments of narrowband and wideband signal microphone array positioning in a real indoor environment were performed.The DOA estimation of the experimental data was achieved using the MUSIC algorithm and the ISSM algorithm,and the spatial filtering of the experimental signals was achieved using LCMV beamforming.Effectively filter interference signals and keep the required directional signals.Performing noise reduction and enhancement processing on the obtained signal can achieve the restoration of the original signal.Thirdly,to address the shortcomings of traditional DOA algorithms that cannot achieve sound signal classification,this paper uses the CNN + LSTM network method to train the network with the signal spectrum and GCC-PHAT spectrum as input features.It realizes the identification and localization of multiple static sound sources coexisting in complex environments.Directional separation of overlapping signals is achieved using known DOA information.Tests show that the performance of the network is significantly better than that of the CNN network alone and the CRNN network using only the spectrum as input features.Finally,Aiming at the problem of identifying and locating moving sound sources,a down-sampling process of the reference real DOA direction trajectory is proposed,which solves the problem that the output DOA data after network pooling does not match the reference trajectory.The directional sound intensity vector of the Ambisonic spatial audio format is introduced as an input feature,which reduces the DOA error and realizes the positioning and identification of single and two moving sound sources.The above theories and methods have been verified by simulation and experimental data tests.The results show that the proposed method for localization and identification of sound sources in complex environments can achieve the classification of multiple sound sources and DOA trajectory tracking,and has a certain ability to resist reverberation and noise.