Evaluation of a method for separating digitized duet signals
Robert C. Maher · Journal of the Audio Engineering Society · 1990
A new digital signal-processing method is presented for separating two monophonic musical voices in a digital recording of a duet. The problem involves time-variant spectral analysis, duet frequency tracking, and composite signal separation. Analysis is performed using a quasi-harmonic sinusoidal representation based on short-time Fourier transform techniques. The performance of this approach is evaluated using real and artificial test signals. Applications include background noise reduction in live recordings, signal restoration, musicology, musique concrete, and digital editing, splicing, or other manipulations. criteria may be identified. If two interfering signals occupy nonoverlapping frequency bands, for example, Separation of superimposed signals is a problem of the separation problem can be solved by using freinterest in audio engineering. For example, it would quency-selective filters. In other cases the competing often be useful to identify and remove undesired in- signals may be described in a statistical sense, allowing terference (such as audience or traffic noise) present separation using correlation or a nonlinear detection during a live recording. Other examples include sep- method. However, most superimposed signals, such aration and replacement of errors in a recorded musical as two musical instruments playing simultaneously, do performance, separation of two simultaneous talkers not allow for such elementary decomposition methods, in a single communications channel, or even adjustment and other strategies applicable for signal separation of the level imbalance occurring when one musician must be discovered. in an ensemble briefly turns away from the microphone. In the case of ensemble music, sounds emanating Considered in this paper is a digital signal-processing from different musical instruments are combined in an approach to one aspect of the ensemble signal separation acoustic signal, which may be recorded via a transducer problem: separation of musical duet recordings. The of some kind. Despite the typical complexity of the primary goal of thisproject was to develop and evaluate recorded ensemble signal, a human listener can usually an automatic signal separation system based primarily identify the instruments playing at a given point in on physical measurements rather than psychoacoustic time. Further, a listener with some musical training or models of human behavior, experience can often reliably transcribe each musical In order to separate the desired and undesired signals voice in terms of standard musical pitch and rhythm. we must resort to prior knowledge of some aspect of Unfortunately the methods and strategies used by human the superimposed signals, whereby a set of separation observers are not introspectable and thus cannot serve easily as models for automatic musical transcription