AUTOMATIC SINGING EVALUATION WITHOUT REFERENCE MELODY USING BI-DENSE NEURAL NETWORK

Automatic singing evaluation without reference melody has long been a difficult problem. This paper aims to pilot a novel data driven approach to tackle this artistic problem. We constructed a large scale dataset and designed an innovative Bi-Dense neural network which can address this task efficiently. Though the singing evaluation is quite a subjective task and depends a lot on listeners' preferences, we showed that a specific group has consistency on the singing evaluations, and it is possible to train a model to learn the subjective preferences of this group. In this paper, a large amount of singing clips and corresponding human gradings were collected. And an elaborate designed Bi-DenseNet was trained to discriminate the good singings from the poor ones. The experiments demonstrated the proposed network performs better than the existing networks for singing evaluation task.