Simon Schwär

The Multi-Scale Spectral (MSS) loss is widely used for comparing audio signals, offering a good balance between temporal and spectral resolution, while allowing for phase differences between waveforms that are perceptually irrelevant. However, the configuration of this loss function, including parameters such as window type and size, hop size, and magnitude compression, is often chosen empirically and without explicitly considering the impact on loss behavior. This is particularly relevant in the context of differentiable digital signal processing (DDSP), where loss gradients are back-propagated through fixed DSP building blocks before they are used as learning signals. In this webinar, the presenter gives an overview of various MSS loss configurations and analyzes the effects of individual loss parameters in detail. Using common DDSP components such as oscillators and filters, they illustrate cases where trade-offs between configuration choices become important. Furthermore, they present examples where the MSS loss fails to provide meaningful gradients entirely and discuss potential workarounds proposed in literature.