Rhythm Tracking Using Multiple Hypotheses
David Rosenthal, Masataka Goto, Yoichi Muraoka · 1994
We briefly describe two rhythm-tracking systems, called, respectively, Machine Rhythm and BTS. Given a MIDI stream as input, Machine Rhythm produces an interpretation that is essentially isomorphic to the rhythmic information represented in normal musical notation. The output of the program defines the placement of measures and assigns rhythmic values (half-note beats, quarter-note beats, etc.) to each note. Although Machine Rhythm is not a real-time system, it processes the MIDI information sequentially, paving the way for possible future real-time implementations. The program attempts to handle some of the more sophisticated rhythm-tracking operations of which humans are capable, such as changes from duple to triple meter or changes in tempo. BTS tracks beats using raw audio signals as input 0 in general, a much more difficult problem than tracking it from MIDI data. BTS accomplishes this task by leveraging the fact that for a large corpus of music 0 rock and pop songs 0 the beat is indicated with some reliability by the bass and snare drums. BTS’s non-reliance on MIDI data enables it to handle a broad range of multimedia applications for which MIDI-based beat-tracking programs cannot be used, and the fact that it works in real time enables its application in a variety of live performance situations. Both Machine Rhythm and BTS use a similar strategy for managing uncertain or noisy input data 0 namely, the strategy of pursuing multiple hypotheses. 1. Rhythm Tracking Issues Rhythm tracking and related psychoacoustical and psychological issues have been treated by a number of researchers from a variety of disciplines ([Allen and Dannenberg,