Real-Time Music Beat Recognition Based On Deep Learning

Music beat recognition is a very important task in the field of music information processing,and it is used in many occasions.However,most of the existing research focuses on offline beat recognition,which cannot be used in situations that require real-time beat recognition(real-time beat recognition means that the music beats are recognized in real-time during the playing of the music,rather than after the music is played),such as real-time stage lighting,real-time robot dance performances and other occasions.This thesis proposes a realtime music beat recognition method based on deep learning,which is studied from the three aspects of music genre,tempo and the above information,to improve the problem of small amount of data and incomplete information in the real-time beat recognition process,and to realize real-time music beat recognition and effect improvement.The main research content of this thesis is as follows:(1)Aiming at the problem of only focusing on the beat itself and ignoring the information closely related to the beat such as genre in real-time beat recognition,this thesis proposes a method G-Beat(Genre-Beat)combined with genre information to increase global information and improve the accuracy of real-time beat recognition.First,the music signal needs to be sliced,and the Mel spectrum feature of the music signal is obtained after preprocessing.The beat activation function of the music is recognized by using the Bidirectional Long Short-Term Memory neural network,and then the Hidden Markov Model is used for post-processing.Finally,the result of music beat recognition is obtained.In order to solve the problem of inaccurate tempo discrimination,the use of music genre information is proposed to improve the misidentified multiple beat tempo problem.The experimental results show that the use of music content information,such as music genre information,in the process of real-time beat recognition can improve the accuracy of beat recognition and improve the inference efficiency.(2)In view of the fact that genre and tempo must be used as prior knowledge to participate in the beat recognition process,resulting in the model only being able to recognize limited genres of music,this thesis proposes a real-time music beat recognition method T-Beat(TempoBeat)combined with tempo to solve the problem of automatic tempo recognition,so as to be able to identify music of any genre,breaking the limitations brought about by genre information as an intermediate process.This thesis chooses the more commonly used Mel spectrum feature as the input signal of the model and uses the Long Short-Term Memory neural network and comb filter to analyze the problem of tempo recognition.In addition,the beat tempo spectrum features of music are analyzed,and the convolutional neural network is used for learning to realize end-to-end beat tempo recognition.In the experimental stage,the confidence level of tempo recognition is analyzed experimentally,and finally the results of the combination of the two tempo recognition methods and the real-time beat recognition process are analyzed and compared.The experimental results show that the beat tempo recognition algorithm based on convolutional neural network can automatically identify the beat tempo of music.When combined with the original real-time beat recognition method,it can not only improve the time efficiency,but also improve the accuracy of beat recognition.(3)A real-time music beat recognition model combined with a caching mechanism--CTBeat(Context-Tempo-Beat)model is proposed to solve the problem that the model cannot access the above information due to slicing the music.In the process of real-time music beat recognition,slicing music causes the model to be limited to the information of the current slice and cannot access the above information.Even if the global information of the current slice is added,such as genre information and tempo information.In the process of fragment recognition,there will still be cases where the intermediate fragment recognition is completely wrong.This thesis proposes a real-time music beat recognition model combined with a cache mechanism--CT-Beat(Context-Tempo-Beat).For continuously arriving music signals,it is still recognized by slices.After reaching a music signal of sufficient duration,the global historical information is used to obtain the pseudo-label after identification,and then gradually use the historical information and global information to update the buffer beat value in the process,so as to achieve the purpose of using the above information.Experimental results show that the caching mechanism helps to solve the problem of continuous segment recognition errors.