Fast and Accurate Song Recognition: an Approach Based on Multi-Index Hashing

Abstract

An activity of wide interest for researchers and companies working in the field of audio signal processing is the capability to automatically recognize in real-time short excerpts of broadcast or played commercial songs. It appears quite difficult to obtain a robust approach able to generate a fast algorithm in order to analyze several audio flows at the same time. In this paper, we compare the results obtained using a specific improvement of an algorithm we recently proposed against several baseline approaches. Specifically, we introduced an approach based on Multi-Index Hashing which permits to improve noticeably speed in fingerprints searching also on very large datasets. Experimental results, performed using the MTG-Jamendo dataset, containing more then 50, 000 songs, show our approach outperform the others jointly considering performance parameters: accuracy, precision and query time.