Humming Audio Spectrum Recognition Based On Deep Learning

Humming recognition is a hot topic in the current music information retrieval research.It aims to identifying the song hummed by user.Because of its friendly interaction and convenience in mobile devices,related technologies have broad application prospects in music retrieval,music recommendation and other fields.Aiming at the problem of automatic recognition of humming audio signal,this paper studies the common methods of deep learning research,applies deep learning method to humming recognition,and combines the traditional audio signal processing method,proposes a recognition method based on depth learning framework,and designs a deep learning framework for humming audio recognition.It can recognize music signals humming.Based on this,a practical and extensible humming audio spectrum recognition system is implemented,and the feasibility of the proposed hum recognition algorithm based on depth learning is verified by experiments.The specific finished work goes as follows:(1)The method of humming audio data processing is studied,including audio filtering,pre accentuation,window segmentation,note starting point detection,audio signal feature representation and so on.The differences,advantages and disadvantages of different methods are compared,and a suitable method is selected for in-depth study and Practice according to the requirements of deep learning data sets.(2)The theoretical basis and common methods of current deep learning research are studied in depth,including the design and training methods of common neural networks,and the principles,advantages and disadvantages of various kinds of neural network structures.Based on this,a convolution recurrent neural network model is designed,which combines the advantages of the convolution neural network in local feature extraction and recurrent neural network in the data characteristics of the inductive sequence,to improve the accuracy of the humming audio signal recognition.(3)Using the open source deep learning platform and tools,experiments and demonstrations of the proposed neural network model are carried out.Through training and testing on the test data set,the model is adjusted repeatedly,and the better model parameters are obtained.The feasibility and effectiveness of the proposed neural network model are verified by the evaluation test on the test data set,and the performance of the model is analyzed and evaluated.(4)Based on the proposed depth learning framework,a humming audio spectrum recognition prototype system is designed and implemented by using the server and mobile development technology,including the audio recognition service of the server,audio recording and audio uploading of the mobile terminal.