Music to MIDI : progress towards the automatic transcription of multi-timbral musical signals into standard MIDI files

This dissertation addresses the problem of automatically transcribing multi-timbral musical signals into a format compatible with the Musical Instrument Digital Interface (MIDI) specification. A major component of this problem is the necessity to separate performed notes according to the instrument of origin. In response to the proposed problem, a complete transcription system for duets between typical monophonic orchestral instruments has been conceived, developed, and implemented using the LabVIEW® graphical programming language from National Instruments, Inc. This system not only has the capability to detect notes and estimate important parameters such as attack time, release time, and amplitude, but also the capability to decide which of the two instuments was the most likely originator. High resolution spectral analysis using an FFT algorithm and complex envelope extraction using quadrature demodulation are used in a combined fashion in a technique termed adaptive quadrature demodulation (AQD). The information resulting from the application of AQD provides a clear picture of how the amplitudes of the partials of a particular note evolve with time. The relative amplitudes of the partials are used to calculate the normalized average spectral distribution (NASD) which is used as a criterion to determine the instrument of origin. The transcription system was tested against recordings of duets between pairs of four instruments, trumpet, clarinet, violin, and piano. The end results of the transcription experiments were standard MIDI files which faithfully reflected the actual performance. In addition the notes were separated into two tracks, one for each instrument, with an overall success rate of greater than 90%.