Dual Attention Network for Pitch Estimation of Monophonic Music

Abstract

The task of pitch estimation is an essential step in many audio signal processing applications. In this paper, we propose a data-driven pitch estimation network, the Dual Attention Network (DA-Net), which processes directly on the time-domain samples of monophonic music. DA-Net includes six Dual Attention Modules (DA-Modules), and each of them includes two kinds of attention: element-wise and channel-wise attention. DA-Net is to perform element attention and channel attention operations on convolution features, which reflects the idea of "symmetry". DA-Modules can model the semantic interdependencies between element-wise and channel-wise features. In the DA-Module, the element-wise attention mechanism is realized by a Convolutional Gated Linear Unit (ConvGLU), and the channel-wise attention mechanism is realized by a Squeeze-and-Excitation (SE) block. We explored three kinds of combination modes (serial mode, parallel mode, and tightly coupled mode) of the element-wise attention and channel-wise attention. Element-wise attention selectively emphasizes useful features by re-weighting the features at all positions. Channel-wise attention can learn to use global information to selectively emphasize the informative feature maps and suppress the less useful ones. Therefore, DA-Net adaptively integrates the local features with their global dependencies. The outputs of DA-Net are fed into a fully connected layer to generate a 360-dimensional vector corresponding to 360 pitches. We trained the proposed network on the iKala and MDB-stem-synth datasets, respectively. According to the experimental results, our proposed dual attention network with tightly coupled mode achieved the best performance.