Onset Event Decoding Exploiting the Rhythmic Structure of Polyphonic Music

In this paper, we propose a rhythmically informed method for onset detection in polyphonic music. Music is highly structured in terms of the temporal regularity underlying onset occurrences and this rhythmic structure can be used to locate sound events. Using a probabilistic formulation, the method integrates information extracted from the audio signal and rhythmic knowledge derived from tempo estimates in order to exploit the temporal expectations associated with rhythm and make musically meaningful event detections. To do so, the system explicitly models note events in terms of the elapsed time between consecutive events and decodes the most likely sequence of onsets that led to the observed audio signal. In this way, the proposed method is able to identify likely time instants for onsets and to successfully exploit the temporal regularity of music. The goal of this work is to define a general framework to be used in combination with any onset detection function and tempo estimator. The method is evaluated using a dataset of music that contains multiple instruments playing at the same time, including singing and different music genres. Results show that the use of rhythmic information improves the commonly used adaptive thresholding onset detection method which only considers local information. It is also shown that the proposed probabilistic framework successfully exploits rhythmic information using different detection functions and tempo estimation algorithms.