Singer Identification Using Convolutional Deep Belief Networks

With the development of Internet technology and the increasing number of multimedia audio, how to effectively retrieve the desired destination, becomes an urgent problem of music information retrieval(MIR) technology. More and more people pay attention to singer identification(SID) which is a sub-task of MIR. For SID, which features to be selected, opinions vary on this point, and how to do intelligent becomes a serious problem to be solved. On the other hand, people are gradually learning neural network from shallow structure to deep structure. Deep learning simulates the deep level of human brain and extracts the desired characteristics by itself. This paper will focus on applying the convolutional deep belief networks(CDBN), which is a deep learning model, to feature extraction of SID task.This paper elaborates the relevant content of MIR and introduces the existing methods, outlines the music features, including short-time energy, short-time average zero crossing rate, spectrum, linear prediction coefficients, linear prediction cepstral coefficient(LPCC), MEL frequency cepstral coefficient(MFCC) and other audio features. The neural network is introduced and gradually into the deep learning theoretical knowledge, then research on the development of the deep learning, the structure model, especially in CDBN model.This paper designs the parameters of the CDBN model, and applies it to feature extraction of SID for the first time, then through parameter optimization of support vector machine(SVM) for classification. In the experiment, it consists of two parts. In the first part, there are no background accompaniment songs for SID experiment, the results show that using the CDBN feature has higher accuracy than using LPCC and MFCC. In the second part, based on the background accompaniment songs for SID experiment, choosing the same features as contrast, the results also show that using CDBN to extract feature has better recognition effect than the traditional single feature. The experiments also indicate that using deep learning in the field of MIR has great research value.