Carleton University - School of Computer Science Honours Project
Fall 2022
Relative Velocity Prediction for Automatic Drum Transcription using Deep Learning
Harman Kang
SCS Honours Project Image
ABSTRACT
The field of Automatic Music Transcription (AMT) is a broad field interested in learning the human ability to aurally transcribe musical sequences. Automatic Drum Transcription (ADT) is a subfield of AMT which focuses on drum sequences, including rhythm, timing, hit recognition, and velocity. Velocity prediction, or the loudness of a note/hit, is an open problem in the field of AMT and is the focus of this paper. Velocity prediction is difficult as many issues can arise when creating an absolute prediction model as the recording conditions play a large part in how velocity is perceived. To mitigate this issue, we propose a new method of predicting velocities through the prediction of relative velocities using peak normalization along with a normalization of input velocities per sequence.