Estimação de Múltiplos Pitches em Áudio Musical Polifônico Utilizando Rede Neural Convolucional

Abstract

Pitch is the auditory perception of the height of a sound, primarily related to its fundamental frequency. Pitch estimation in musical audio is a fundamental yet challenging task in the field of music information retrieval (MIR). Accurately determining the fundamental frequency (F0) of musical notes is crucial for various applications, including music transcription, melody extraction, and audio analysis. The problem is complicated by factors such as polyphony, background noise, and variations in timbre and dynamics. Traditional methods for estimating pitch, when dealing with these complexities, often produce inaccurate or unreliable results. Recent improvements in deep learning, especially using convolutional neural networks (CNNs), have shown potential in tackling these challenges. CNNs are capable of learning hierarchical representations from raw audio data, effectively capturing the temporal and spectral features essential for accurate pitch estimation. This work explores the problem of multi-pitch estimation in musical audio, highlighting the inherent challenges and the current state-of-the-art solutions using CNNs. We discussed the innovations in the CREPE model architecture and the training strategies that enable the model to estimate multiple pitches. The results demonstrated robust performance in different scenarios. In validation sets with multiple frequencies, the newly proposed model achieved high RPA values, with an overall average close to 0,93, highlighting its ability to correctly identify simultaneous pitches.