The Research Of Music Emotion Classification Based On Audio Signal Processing

From ancient to modern times, music has always been an indispensable part of people's life, with the development of science and technology, digital music has become more and more important. More people choose to search and listen to music from the Internet. Music is the carrier of the emotion, people use emotion as the basis for retrieval of music. Therefore, the research on Music Emotion Classification (MEC) becomes very important, which has a great effect on the development of the field of Music Information Retrieval (MIR).Based on Audio Signal Processing (ASP) Technology, this paper establishes relation between music emotion and the basic elements of music and deduces the calculation of the relevant audio features. Machine learning algorithm is adopted to carry out MEC task and a double-layers MEC system is designed in this paper. The mainly works showed as follows:First, the characteristics of the time and frequency domain of music signal are analyzed. In terms of audio feature extraction, this paper designs and implements the chroma vector on basis of Constant-Q Transform to enhance the accuracy and proposes an algorithm to calculate tonality feature using cross-correlation operation. Also this paper deduces the calculation of beat feature for noise music, selects Mel Frequency Cepstrum Coefficient for the analysis of music timbre.Second, this paper uses mathematical statistics method to integrate long-time and short-time audio features to carry out model training and compare the performance of different models.Third, to enhance the performance, this paper proposes a double-layers MEC system.The first layer of the system aims to carry out rough classification using short-time features.Singular-point rectified algorithm is proposed to enhance the accuracy rate in this part. The second layer is used for fine classification using long-time features. Meanwhile, this part offers reference category for the music which is difficult to distinguish. Furthermore,experiment is designed to compare the double-layers model with other classification models, it proves the double-layer MEC model has a 3.6% better performance than the best single-layer model.