DDSP Study: Magnitude Compression in Spectral Losses Helps Only If It Matches the Informative Region
kastnerkyle · x · 2026-10-20
- A paper submitted to ICASSP 2027 examines whether magnitude compression (e.g., log-magnitude terms) in DDSP spectral losses actually aids optimization.
- The authors introduce the concept of a system's "informative region" — the part of its dynamic range carrying information toward the target — and propose concordance, a metric that locates this region without training.
- Experiments fitting a modal plate reverb, a subtractive synthesizer, and a low-energy-region system show the loss matching the informative region always wins: uncompressed when the region is at high energies, compressed when low.
- Surprisingly, the common log + linear magnitude combination was best in none of the systems.
- Code and audio examples are open-sourced.
More from Multimodal
- Anthropic Dev Shares the Prompts Behind His Opus 5.5 Code-Rendered Films — dfeinition · 2026-10-20
- Designer shares 16-color pairing guide and full Midjourney prompt for editorial still life — tisch_eins · 2026-10-20
- Flux 3 Max I2V test: a full prompt for generating a photoreal horror forest opening shot — umesh_ai · 2026-10-20
- Stylized animation made with Flux 3 Max on fal, then upscaled — BLUECOW009 · 2026-10-20
- Matt Shumer marvels at AI's rapid 3D generation leap in under a year — mattshumer_ · 2026-10-20
- A geospatial memory palace lets you walk through summer and winter of the same place — bilawalsidhu · 2026-10-20