Training Better Embedding With Perturbed Data Augmentation for Automatic Singing Quality Assessment

Abstract

Automatic singing quality assessment is a challenging task due to its subjective nature, as well as limitations stemming from the scarcity of human annotations and the poor generalization to unfamiliar singing expressions. In this paper, we incorporate perturbed augmented data strategies into the training of a singing quality assessment (SQA) network without additional annotations. In addition, we also introduce the training of a classifier to enhance the network’s latent space in response to pitch, tempo, and timbre variation. The experiments demonstrated that our proposed SQA network outperformed the baseline with a 6.67% improvement in terms of the Pearson correlation coefficient.