A post-processing of onset detection based on verification with neural network

Abstract

Onset detection is the primary task of music transcription that aims to find the start time of each note, which directly associated with the beats perception in the auditory system. Researchers attempted to find a data representation of universal onset function. However, the onset detection would not generalize to all cases. For example, onset detection in solo singing has a lower performance than solo playing the instrument in MIREX challenge every year. This paper presents a post-processing step to singing onset detection that solely reduces false detected onsets. In the post-processing step, the system checks the onsets picked from local maximums of onset function, and uses the neural network model to discern onset or non-onset feature rather than consider a complicated onset function. The performance of the network has a close relationship to the onset detection. In the public dataset about the research of singing transcription, the pipeline with post-processing presents a higher performance than the standard and novelty method, when it was focused on the onsets, that it reduces false alarms from feature methods. It can provide further supports for the research of singing transcription when the data-driven approach provided an effective method to eliminate spurious peaks, which can be the state-of-art of singing onset detection.