An EB-enhanced CNN Model for Piano Music Transcription

Abstract

Automatic Music Transcription (AMT) is an important task in Music Information Retrieval (MIR). Many researchers have focused on the structure of Convolutional Neural Network (CNN) for transcription. In this paper, we construct a CNN-based (EB) piano music transcription model using the energy-balanced constant Q transform spectrogram, which is called EB-enhanced CNN model. Unlike standard CNN-based methods, the proposed model makes the energy of the input features more balanced, so that many previously missed pitches due to weak energy can be successfully detected. Training and evaluation are performed on the MAPS dataset, a public dataset for piano transcription. As a result, our technique achieves a 3.53% f1 score improvement compared with the state-of-the-art method on the MAPS ENSTDkCl subset.