Identificação de acordes de violão utilizando machine learning

Abstract

Music transcription is the act of annotating a sound or piece generated by an instrument that has not been previously annotated. It is a manual work that requires knowledge in music theory. Chord annotation task is part of musical transcription. This task can be done using machine learning. Studies to create automatic chord annotation tools using machine learning have focused on the task of classification and post-processing. Recently, it has been shown that focusing on pre-processing, with algorithms that make dataset examples more descriptive about the chord, can yield better results. The work shows the creation of a dataset using an efficient method for extracting features from 41-chord recorded audio. After construction of the dataset, models based on logistic regression, neural networks and random forests were tested. Among all the models tested, the random forest obtained the best results, combined with a feature extraction algorithm that uses the Constant-Q transform technique. The model’s average accuracy reached 99%.