An Approach for the Separation of Voices in Composite Musical Signals

The simultaneous presentation of several sound sources in a performance setting is fundamental to most music. Ensemble musical signals consist of superpositions of multiple distinct sonic events which mayor may not be synchronized in time, frequency, and/or amplitude. Once the distinct events are combined and recorded in a storage medium, e.g., a digital recording, the composite signal is often unsatisfactory in some way: the recording might suffer from poor ensemble balance, performance errors, or corruption from undesired background audience noises (sneezing, talking, etc.). Although it often might be helpful to process the constituent signals independently, separating the composite signal into its parts is a nontrivial task. The research reported here considers particular aspects of the separation problem: analysis, identification, tracking, and resynthesis of a specified voice from a digital recording of a musical duet setting. Analysis is accomplished out of real-time using a quasi-harmonic, sinusoidal representation of the constituent signals, based on short-time Fourier transform (STFT) methods. The procedure is evaluated via resynthesis of a "desired " signal from the composite analysis and tracking data. Other applications include signal restoration, digital editing and splicing, musique concrete, noise reduction, and time-scale compression/expansion. iii This material is based upon work supported, in part, under a National Science Foundation Graduate Fellowship. Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author and d