Audio Feature Extraction: Research on Retrieval and Matching of Hummed Melodies

Abstract

Humming clips facilitate more intuitive and user-friendly music retrieval. This paper combined the K-means clustering algorithm, back-propagation neural network (BPNN) model, and dynamic time warping (DTW) algorithm for humming music retrieval and matching.  Initially, the K-means algorithm was used to narrow down the search scope. Then, the BPNN model was employed to extract the abstract features of the music melody, and the DTW algorithm was used to match these abstract features. In the simulation experiment, the classification ability of the K-means algorithm was verified, and then it was compared with the DTW and BPNN+DTW retrieval algorithms. The results showed that the K-means algorithm had good music segment classification performance. The retrieval algorithms used could retrieve the target music more accurately and stably.