Research On Singing Voice Separation Of Mono Music

Monaural singing voice separation is both the focus and the difficulty of content-based music analysis.This paper analyzes and improves two most advanced algorithms.The main contents and innovations of this thesis are as follows:1.It summarizes the current monaural singing voice separation methods from the perspectives of accompaniment and vocal,and divides it into two types based on the accompaniment repetition pattern and the vocal pitch estimation.In the review of the above two kinds of methods,the main ideas and the existing constraints of the various methods are listed,which provides a preliminary direction for the proposed modifications.2.This thesis chooses the nearest neighbor median filter method as the referenced method,which is not suitable for dealing with mixture signal with strong vocal energy.An improved singing voice separation method is then proposed with the multiple sub-bands partition combined with the nearest neighbor median filtering.When the vocal energy of mixture signal is strong,the partitioning of multiple sub-bands makes accompaniment energy of some sub-bands dominate,which in turn improves the separation performance.Both the referenced and proposed methods are evaluated objectively with the mainstream test dataset MIR-1K.The results show that the proposed method is better than the referenced method,and the overall evaluation factor of separated vocal is improved by at least 2.5dB meanwhile the overall evaluation index of separated accompaniment is increased by at least 0.8dB.3.This thesis chooses instantaneous mixture model(IMM)method as the other referenced method and finds vocal pitch estimation method is too simple and gives inaccurate result.After that,Melodia and vocal detection algorithms are introduced to the performance improvement.The 2nd pair of referenced and improved methods are evaluated objectively on test dataset MIR-1K again.The results show that the separation performance is indeed improved by introducing Melodia algorithm,the overall quality of the separated vocal improves at least 1.67dB,in the same time the overall evaluation factor of separated accompaniment increases at least 1dB.After including vocal detection,the separated vocal contains less accompaniment component,but due to the vocal detection performance constraints,performance improvement can only be seen in the-5dB test dataset.The objective evaluation results of the test dataset show effectiveness of the proposed methods,and rationality of the discussion and analysis of the referenced methods.