A Novel Perceptual Loss Function for Audio and Music Quality Differentiable in PyTorch

This paper presents a novel perceptual loss function designed for evaluating audio and music quality, implemented in PyTorch and available via GitHub and Google Colab. The loss function integrates psychoacoustic modeling by operating in the Modified Discrete Cosine Transform (MDCT) domain and compares the difference between the original and modified audio signals against a psychoacoustic masking threshold. The function's behavior as a divergence rather than a distance metric, along with its differentiable properties, makes it a useful tool for fine-tuning deep learning models in audio signal processing. We provide an in-depth discussion of its implementation, as well as a comparison with traditional and learned loss functions. This function offers an advantage over existing methods due to its psychoacoustic basis, resulting in a more perceptually accurate measure of audio quality.