Interpreting Rhythm in Optical Music Recognition

We present a method for understanding the rhythmic content of a collection of identified symbols in optical music recognition, designed for polyphonic music. Our object of study is a measure of music symbols. Our model explains the symbols as a collection of voices, while the number of voices is variable throughout a measure. We introduce a dynamic programming framework that identifies the best-scoring interpretation subject to the constraint that each voice accounts for the musical time indicated by the known time signature. Our approach applies as well to the situation in which their are multiple possible hypotheses for each symbol, and thus combines interpretation with recognition in a top-down manner. We present experiments demonstrating a nearly 4-fold decrease in the number of false positive symbols with monophonic music, identify missing tuplets, and show preliminary results with polyphonic music.