Vocal Melody Extraction Based on Sparse Autoencoding Neural Networks

In music, melody is an essential element. Extracting the main melody is a key technique in music search. In this study, we propose a new algorithm for automatically extracting the vocal main melody. Using a more efficient sparse autoencoding neural network in place of the original shallow BP neural network can improve the recognition accuracy of the primary melody model, reduce the false alarm rate of melody localization, and thus enhance the overall accuracy of vocal melody extraction. Empirical testing on the MIR-IK dataset proves that the improved algorithm can significantly increase the accuracy of vocal melody extraction, increasing more than 1.51% and reducing the average extraction time by 0.12 s.