Generative audio models based on neural networks have led to considerable improvements across fields including speech enhancement, source separation, and text-to-speech synthesis. These systems are typically trained in a supervised fashion using simple element-wise l1 or l2 losses. However, because they do not capture properties of the human auditory system, such losses encourage modeling perceptually meaningless aspects of the output, wasting capacity and limiting performance. Additionally, while adversarial models have been employed to encourage outputs that are statistically indistinguishable from ground truth and have resulted in improvements in this regard, such losses do not need to explicitly model perception as their task; furthermore, training adversarial networks remains an unstable and slow process.
In this work, we investigate an idea fundamentally rooted in psychoacoustics. We train a neural network to emulate an MP3 codec as a differentiable function. Feeding the output of a generative model through this MP3 function, we remove signal components that are perceptually irrelevant before computing a loss. To further stabilize gradient propagation, we employ intermediate layer outputs to define our loss, as found useful in image domain methods. Our experiments using an autoencoding task show an improvement over standard losses in listening tests, indicating the potential of psychoacoustically motivated models for audio generation.
A summary of the work and sound examples can be found here: https://ishwaryaanant.github.io/small-network-perceptual-loss/demo