A Multi-Modal Joint Voice Parts Division Method Based On Deep Learning

The division of singers' voice parts is a basic problem in vocal arts medicine and vocal music teaching. Improper voice parts classification can have a negative impact on singers' health. In this work, we propose an objective method to classify the voice parts from the acoustic perspective and anatomic perspective. We first extract relevant features from singers' voice signals and laryngoscope images, then splice the features of these two modalities. After that, a voice parts division convolutional network (VDCNN) model is constructed to evaluate singers' voice categories. In order to realize this model, real data of voice signals and laryngeal images from 74 singers is applied to train and adjust this model. The analysis of experimental results shows that the characteristic parameters and network structure used in this model are significantly improved compared with similar methods, and the optimal classification accuracy is 93.17%.