Intelligent Music Chorus Recognition and Evaluation System Based on Deep Learning

Abstract

The key to the chorus recognition is to find the effective representation of the chorus timbre. Through the analysis of the temporal frequency domain features, inverted spectrum features, sparse features and probability features, it shows that the chorus can be effectively identified by using the temporal and frequency domain features. Based on this result, using the layer-by-layer abstraction feature of deep learning, the advanced time-frequency representation of the chorus timbre is extracted for the chorus recognition. This paper aims to study the design of an intelligent music chorus recognition and evaluation system based on deep learning. Considering the problem of high classification error, using the time domain frequency domain characteristics, clutter characteristics, sparse characteristics and probability features to identify the adverse effects of impact instrument, put forward a low noise instrument recognition model is a use of cochlear model harmonious decomposition music, including time and frequency information auditory spectrum, similar to the human hearing. In order to combine the feature expression ability of the serial-level noise cancellation encoder and the abstract feature ability of the deep belief network, the five-layer deep hybrid network is constructed for the deep learning framework with the above two basic modules. It is proved that the evaluation error of chorus recognition and professional judges is less than 5%. And it greatly improves the evaluation efficiency.