Advancing Transposition-Preserving Pitch Estimation in Audio Signals with Neural Network Architectures

This study addresses the challenge of transposition-preserving pitch estimation from audio signals, a critical task in music information retrieval that facilitates robust pitch recognition across varied musical transpositions. Accurate pitch estimation is foundational for applications like automated music transcription and music analysis, where maintaining the relational integrity of pitch shifts is essential. The objective of this research is to develop a neural network model capable of predicting pitch distributions from Constant-Q Transform (CQT) frames that are both original and pitch-shifted. The proposed model incorporates advanced neural architecture involving convolutional layers and transposition-preserving Toeplitz fully connected layers. Specifically, the model processes input through layer normalization and a series of 1D convolutions, integrating Squeeze-and-Excitation (SE), Convolutional Block Attention Module (CBAM), and Coordinate Attention (CA) to enhance feature recognition capabilities. This is followed by a softmax layer for classification, ensuring the model's outputs reflect the transposed relationships evident in the inputs. Experiments conducted on a comprehensive audio dataset demonstrate that the integration of attention mechanisms significantly enhances model performance, with the Coordinate Attention module proving particularly effective in spatial feature recognition. The results highlight the capability to preserve pitch information across transformations and confirm its potential in real-world applications. This study not only advances the field of music information retrieval but also establishes a framework for future developments in pitch estimation technologies.