A rhythm practice support system with annotation-free real-time onset detection

In this paper, we discuss a system for musicians to practice rhythm with a musical instrument. Our system offers real-time visual feedback to the user about the rhythm played along with guiding metronome sounds. The system uses spectral flux and dynamic thresholding to find onsets in the audio signal. Usually, a real-time onset detection method uses annotation data, which need to be set manually and require a great deal of effort to make, to choose optimal parameter values for the thresholding. We propose a method to calibrate parameter values using sample sounds played in advance by the user instead of using annotation data. This makes our system much practical to be used. We conducted experiments with wind instrument players using our system and the results showed it helps the users improve the rhythm played.