DAFx paper cuts reverb modal estimation error by 66% with hybrid ML

Estimating the parameters of a plate-reverb impulse response—frequencies, decay rates, gains, and mode count—is a notoriously hard task when modes are weak and overlapping. Sparse peak detection often undercounts these modes, skewing the entire fit. To solve this, Minhui Lu and Joshua D. Reiss from Queen Mary University of London trained an ExtraTrees regressor on simulator-generated data to predict the number of modes in four frequency bands. These predicted counts define dense frequency grids, and a differentiable all-pole resonator model then refines decay and gain while keeping frequencies fixed.

The system was tested on two synthetic validation sets from the 1st DAFx Parameter Estimation Challenge (Task B). It reduces a local challenge-style error by about 66% relative to the official default peak-picking baseline. The improvement largely stems from lower mode-count mismatch, while decay and gain estimation still remain the largest error contributors. The authors argue that separating modal-density estimation from continuous parameter fitting is a more robust strategy than trying to do both simultaneously, and the approach could generalize to other dense modal analysis problems beyond reverb.