Automatic Drum Transcription (ADT) using ConvNets

Recent efforts in Music Information Retrieval (MIR) have improved digital processing algorithms for getting accurate high-level information about music signals. Those algorithms have led to the development of musical signal processing applications; for instance, the detection and the extraction of percussive events in a complex drum track of a professional drummer, also the classification of these unpitched sounds related to percussion instruments, such as the snare drum (SD), the kick or bass drum (KD/BD) and the associated cymbals, like the hi-hat (HH), as Figure 1 reveals. Where these analysis tools help to generate a transcription of the input drum sample which corresponds to an output and describes this audio signal by the onset times and its percussive sound category, and this MIR task is known as ADT.

Figure 1: Main percussion instruments that conform a basic drum kit for ADT systems [1].

The rising enthusiasm in ADT research has increased, due to the availability of Artificial Intelligence (AI) technologies over multiple platforms, using Machine Learning (ML) and Deep Learning (DL) algorithms, also the number of available datasets has increased. Therefore, the following article describes the implementation and the experimental results of an ADT system based on two stages: an onset detection-extraction scheme and a DL classifier to generate high-level information about an audio sample that consists of a drum track with several percussive…