Data Generation from Robotic Performer for Chord Recognition

摘要

This paper presents a synthetic data generation method using a robot to create a substantial dataset. One important task in the field of learning-based recognition is to collect large amounts of high-quality training data. To increase the training dataset, many researches have used data augmentation methods. In musical recognition, data augmentation is implemented using digital signal processing methods including pitch-shifting and time-stretching. Data augmentation is a limited method because it depends on prior knowledge of the data and it cannot be performed all domains. We propose a new dataset collection method using a robot that automatically plays musical instruments, which enables high-quality data to be added to the training samples. We compare the results with two kinds of human dataset and a mixed dataset, which include human and robot datasets, using four kinds of convolutional neural networks (CNNs). The results indicate that the proposed method using CNNs analyzing the mixed dataset with a guitar-playing robot, can outperform CNNs using the human dataset.