Unsupervised training of detection threshold for polyphonic musical note tracking based on event periodicity

A common approach to the detection of simultaneous musical notes in an acoustic recording involves defining a function that yields activation levels for each candidate musical note over time. These levels tend to be high when the note is active and low when it is not. Therefore, by applying a simple threshold decision process, it is possible to decide whether each note is active or not at a given time. Such a threshold, in general, is hard to set and has no physical meaning. In this paper, it is shown that the rhythmic characteristic of the musical signal may be used to obtain a suitable threshold. The proposed method for obtaining the threshold is shown to have a greater generalization capability over different databases.