Loss compression choice improves ddsp parameter fitting accuracy
Rethinking Magnitude Compression in Spectral Losses for DDSP Parameter Estimation
Sound
Summary
When teaching computers to recreate sounds, it helps to measure how close the computer's guess is to the real sound. People often shrink (compress) parts of the sound to match human hearing better when measuring this. The authors show that whether this shrinking helps depends on where the important sound information lies in the loudness range. They created a new way to test which measurement works best without training the computer first. Their tests show using the right kind of shrinking for the sound's key parts works better than mixing methods.
What this means in practice
- •For audio software developers: Choose appropriate loss functions to improve parameter estimation accuracy when modeling audio effects and synthesizers automatically.
- •For sound designers: Use better fitting methods to replicate or modify acoustic effects and synthesized sounds more precisely based on target audio.
Authors
Byunghoo Park, Jayeon Yi, Minje Kim
Abstract
Spectral losses for differentiable digital signal processing (DDSP) commonly compress magnitudes, for example by using log-magnitude terms. Such compression mimics human perception, but whether it helps the optimization is less clear, and is usually found only by comparing trained models. In this work, we show that whether compression helps depends on the system's informative region: the region of its dynamic range that carries the information pointing toward the target. To indicate where this region lies without having to train, we introduce concordance, which measures how well a loss orders candidate parameter settings by their distance from the target. We then compare uncompressed, compressed, and combined losses by fitting a modal plate reverb and a subtractive synthesizer model to target sounds. A third system, selected for its low-energy informative region, tests the relationship in the opposite direction. In every system, the loss that matches the informative region performs best: the uncompressed one where the region lies at high energies, and the compressed one where it lies low. Surprisingly, the common combination of both log-magnitude and linear-magnitude terms performs best in none of the systems. Code and audio examples are available at https://github.com/superstem000/compression-analysis