The Classification and Judgment of Abnormal Problems in Music Song Interpretation Based on Deep Learning

Abstract

Song singing interpretation is one of the people’s favorite entertainment pastimes, but there are some abnormal problems in the process of song singing. Traditional song singing problems need to be analyzed and judged by professionals, but the traditional way requires offline face-to-face teaching and is time-consuming and laborious. In this paper, we hope to realize the automatic judgment classification of abnormal vocal problems of song singing and provide some guidance help to online teaching. In this paper, we propose a deep learning-based method for classifying abnormal vocal interpretation problems in music songs, using a computer to record the singers’ voices, and then analyzing and judging them with a trained method model to point out the main problems that exist in the process of singing songs. In this paper, more than 300 singers’ audio were collected and the data were calibrated and classified by researchers specialized in the music field into seven main categories. Short-time Fourier transform (STFT), Mel frequency cepstrum coefficient (MFCC) and spectral mass center methods were used to extract the features of song audio and produce the corresponding datasets. The dataset is trained using residual neural network and EfficientNet. The experimental results of the model in this paper show that the data training accuracy is about 90.1%, which achieves a good result.

Keywords

• ResNet
• EfficientNet
• the spectral center of mass
• deep learning
• short time fourier transform
• mel frequency cepstrum coefficient