A Real-time Music Scene Description System: Detecting Melody and Bass Lines in Audio Signals

This paper describes a predominant-pitch estimation method that enables us to build a realtime system detecting melody and bass lines as a subsystem of our music scene description system. The purpose of this study is to build such a real-time system that is practical from the engineering viewpoint, that gives suggestions to the modeling of music understanding, and that is useful in various applications. Most previous pitch-estimation methods premised either a single-pitch sound with aperiodic noises or a few musical instruments and had great difficulty dealing with complex audio signals sampled from compact discs, especially discs recording jazz or popular music with drum-sounds. Our method can estimate the most predominant fundamental frequency (F0) in such signals containing sounds of various instruments because it does not rely on the F0’s frequency component, which is often overlapped by other sounds’ components, and instead estimates the F0 by using the Expectation-Maximization algorithm on the basis of harmonics’ frequency components within an intentionally limited frequency range. It also uses a multiple-agent architecture to stably track the temporal trajectory of the F0. Experimental results show that the system is robust enough to estimate the predominant F0s of the melody and bass lines in real-world audio signals.