Research On The Separation Algorithm Of Instrumental Music Based On Deep Learning

Music separation is a branch of audio separation task,while instrumental separation is the expansion of music separation task in depth and width.The first explanation of instrumental music separation is based on music separation,which has the same difficulty as audio separation task.The purpose is to separate instrumental audio from mixed audio,and the separated specific audio can be used for many subsequent tasks.The second explanation is to separate specific instrument playing skills from existing audio,and the separated skill audio can be well applied to musicians' performance learning and music analysis.At present,music separation algorithms based on deep learning mostly separate songs,but instrumental music as an important part of music,especially national instrumental music,has not received due attention.In order to alleviate this situation,this dissertation makes a dataset Breath based on bamboo flute audio recording for use,which includes bamboo flute music separation set,bamboo flute skill separation set and bamboo flute sound separation set.The Bamboo flute music separation set is used for the sound source separation task of bamboo flute audio extraction,and bamboo flute skill separation set is used for bamboo flute skill classification task.Aiming at the first explanation of instrumental music separation,this dissertation designs a RECA module based on residual and attention mechanism,which improves the original U-net model for music separation;compared with its two variants and other deep learning algorithms.Compared with its two variants and other deep learning algorithms,the performance evaluation index of RECA-U-net model trained with MUSDB18 dataset for four track separation task is much better.RECA-U-net also has advantages over its two variants in the task of voice song separation.The RECA-U-net model is trained by Bamboo flute music separation set of Breath,and the separation performance of the model is evaluated by subjective evaluation criteria,and the application scope of the model in Bamboo flute music is given.Aiming at the second explanation of instrumental music separation,this dissertation constructs the instrumental music reference models of Breath1 d and Breath2 d based on the Bamboo flute skill separation set of Breath.In the second classification of the divided subsets,Breath1 d and Breath2 d models are seperatedly used to distinguish the Bamboo flute skills,and the best model suitable for the classification of different skills is obtained.In the classification experiment of the complete set,the performance of skill separation is improved to 91.3% by using the data enhancement method and fusing the Breath1 d and Breath2 d models.The completeness of the dataset is proved by analyzing the reasons for the performance improvement.