Single-channel Aliased Bird Sound Separation Based On Attention Mechanism And Empirical Mode Decomposition

Bird species monitoring is an important part of the study of bird population biology,and bird diversity monitoring is the focus of biodiversity monitoring.Because of the long-distance transmission,easy collection,and uniqueness among different species,birdsong is an important basis for bird identification.Therefore,the intelligent recognition system based on birdsong replaces the traditional methods of bird species monitoring,such as strop transects,spot mapping,and so on.However,bird classification based on birdsong is a difficult task.For example,the recoding has relatively large background noise,multi-label classification problems.Therefore,it is necessary to separate the sound source of mixed birdsong.This paper mainly studies the single-channel sound source separation method based on deep neural networks and the birdsong recognition method based on deep neural networks and designs a birdsong system from separation to recognition.The main contents of this paper are as follows:Firstly,this paper studies the application of single channel speech separation method based on deep neural network in mixed birdsong.In order to further improve the separation performance,a single channel birdsong separation method based on attention mechanism is proposed.Attention mechanism helps the model focus on more meaningful information and improves the separation performance.Our experiments on the birdsong dataset results in13.6d B of the scale-invariant signal-to-noise ratio.Secondly,in view of the high overlap between the time domain and frequency domain of the mixed birdsong,the ensemble empirical mode decomposition(EMD)is added to the single channel birdsong separation method based on attention mechanism,and the decomposed eigenmode component is used as the input of the neural network instead of the mixed birdsong.The experimental results show that the scale-invariant signal-to-noise ratio reaches 16.8d B,and it has a better separation effect than the current mainstream separation algorithm in the case of two sound sources.Finally,this paper studies the birdsong recognition method based on a deep neural network and proposes a birdsong recognition method combining convolutional neural networks and recurrent neural networks.In order to solve the problem of imbalanced categories of the birdsong dataset,the data is enhanced and the loss function is modified.Experimental results show that the average recognition rate of the model reaches 83.1%,the data enhancement and loss function improvement can effectively improve the average recognition rate and robustness of the model.At last,it is verified that the processing of the single-channel birdsong separation system can improve the average recognition rate of the birdsong recognition system.