Back to Ear: Perceptually Driven High Fidelity Music Reconstruction

ABSTRACT V ariational Autoencoders (V AEs) are essential for large-scale audio tasks like diffusion-based generation. To address these challenges, we propose ϵar-V AE, an open-source music signal reconstruction model that rethinks and optimizes the V AE training paradigm. Our contributions are threefold: (i) A K-weighting perceptual filter applied prior to loss calculation to align the objective with auditory perception. Experiments show ϵar-V AE at 44.1kHz substantially outperforms leading open-source models across diverse metrics, showing particular strength in reconstructing high-frequency harmonics and the spatial characteristics. Index T erms-- V AE, Music, Phase, Perceptual Weighting 1. INTRODUCTION Achieving perfect, perceptually lossless reconstruction of complex audio signals like music remains a central challenge in audio engineering and machine learning.