PHAST-Net: Attention-Guided, Physics-Informed Network for Estimation of Ideal Time-Frequency Representations
PHAST-Net is an attention-guided network designed to estimate Ideal Time-Frequency Representations (ITFRs) across various domains like spectrograms and tempograms. It uses a set of wavelet transforms and a physics-informed loss function to learn mappings from input signals to high-resolution, cross-term-suppressed representations. This allows for robust analysis of speech and music.