Abstract: In the field of music information retrieval,the extraction of the main melody is a very difficult task.In the polyphonic music,different sound sources interact with each other,leading to discontinuity of the main melody’s pitch sequence,which reduces the accuracy of the original pitch of the melody.In response to this problem,a CNN-CRF model with enhanced pitch salie-ncy representation and automatic melody tracking is designed.In order to better extract the harmonic information,it is proposed to enhance the initial saliency representation of the SF-NMF calculation by structured data,and to combine the melody characte-ristics and the smooth constraint conditions of the pitch under the dynamic programming framework to find the optimal evolution path.Experiments show that the proposed method has better melody extraction results,and the original pitch accuracy on both test data sets is higher than that of other reference methods.Comparing different inputs validates that structured data can enhance the significance representation and make up for the misJudgement of pitch by SF-NMF.

Key words: CNN-CRF, Melody extraction, Music information retrieval, Music signal processing, Pitch saliency enhancement