The Study Of Automatic Notes Segmentation Based On DPP

In the era of data explosion,how to classify and manage the mass media information has become a difficult problem.The traditional method of manual annotation is far from meeting the needs,so the music element analysis based on content has become a hot research topic.Content based music analysis is an important branch of the field of intelligent processing of computer music,which is one of the key techniques in the segmentation and recognition of musical notes.At present,it has been a reliable algorithm to convert each period of time signal to the audio signal,but there is no good algorithm can be converted to the pitch sequence to obtain the discrete note sequences for accurate segmentation,automatic recognition of the notes is still a technical problem.On the basis of the results of the research on the segmentation of the notes,the dissertation proposes a new algorithm for the segmentation of the notes based on the theory of music,pattern recognition and machine learning.In this dissertation,the application background,basic concepts,mathematical logic and geometric significance of the four aspects of the determinant point process of a comprehensive analysis,the dissertation expounds the feasibility of the determinant point process,for the follow-up work to provide theoretical support.The determinant point process model is a seed set selection model.Firstly,a 12 dimensional feature vector is established for each frame.The model is trained with the principle of the supervised learning.Finally,a sampling algorithm is used to select a subset of the DPP distribution.Specific work is as follows:First,the music data is pre-processed,the music is unified for the 11025 Hz sampling rate of 3 seconds of music clips,and removed the continuous repetition of the fragment.By means of a sub frame,the continuous music signal is abstracted as a discrete point process.After analyzing and deriving the principle of PCP feature extraction,a 12 dimensional PCP feature vector is built for each frame.Second,a music clip is manually annotated with a note frame number corresponding to the table,and is used for the manual pick a frame subset for training,which is composed of a frame of each note.According to the maximum likelihood estimation(MLE)principle,the objective function is set up,and the objective function can be transformed into convex optimization problem,and the gradient descent method is used to solve the problem.Finally,the DPP sampling algorithm is used to extract the frame of the test data,and the statistical error rate is compared with that of the control table.In this dissertation,the experimental results of the cross validation of the 200 segment of the experimental data,the final segmentation accuracy of 67.92%,different from the traditional signal processing methods,to provide a new method for the notes.