Transcrição automática de música: explorando o treinamento de algoritmos

Abstract

This study addresses AMT, a technique that converts musical audio into written notation such as scores. Initially, it validates the system proposed by Maman e Bermano (2022), which combines supervised and unsupervised methods for musical data labeling and training. Next, it develops a modular system focused on two core areas of AMT: pitch recognition and instrument recognition. To achieve this, a dataset of audio segments labeled with these categories is organized and used to train a CNN. Experimental results show F1-scores above 0.90 in certain classes, while minority classes face limitations due to data imbalance. The model is evaluated through validation metrics, demonstrating its effectiveness in specific recognition tasks. However, in high-polyphony scenarios, the system encounters challenges that indicate the need for future investigations, such as the adoption of recurrent networks or attention mechanisms. This study contributes to the advancement of AMT techniques by providing a methodological foundation that supports research aimed at improving musical transcription.