Toward Automatic Sound Source Recognition: Identifying Musical Instruments

One of the broad goals of research in computational auditory scene analysis (CASA) is to create computer systems that can learn to recognize sound sources in a complex auditory environment. In this paper, a set of acoustic features is proposed that relate to the physical properties of sound-producing objects. In particular, a set of orchestral musical instrument sounds is presented as representative of the class of sounds produced by quasi-periodic excitation of resonant structures, acoustic properties of this class are considered, and the log-lag correlogram is presented as a signal representation that codes many of the proposed features. Specific examples are given of features extracted from violin, trumpet, and flute tones. Extensions to Ellis’s prediction-driven CASA framework are proposed in the form of a hierarchy of sound-source models represented by frames. It is suggested that the goal of building an artificial system for sound source recognition in complex mixtures may be well served by such an approach.