A new preprint tackles Task A of the first DAFx Parameter Estimation Challenge: estimating six effective physical parameters of a synthetic plate-reverb model from a single impulse response. The authors propose two Transformer-based pipelines, both built around an Audio Spectrogram Transformer encoder.
The first method, A1, feeds encoder features to a Transformer regressor and then refines the output with a differentiable model of the plate-reverb. The second, A2, uses the same encoder to condition a continuous normalizing flow, then refines sampled candidates with particle swarm optimisation and gradient polishing. Both approaches preserve the absolute scale of the impulse response, which is needed to recover surface density.
On a synthetic holdout set of 100 impulse responses, both refinement procedures reduced waveform and parameter errors by more than three orders of magnitude compared with the unrefined neural outputs. The particle-swarm pipeline achieved the lowest errors, approaching perfect recovery in this matched synthetic setting.
The reported near-perfect results are limited to a matched synthetic setting, where the test impulse responses come from the same model used for training. Still, the work demonstrates that combining learned estimators with explicit refinement can make parameter recovery highly accurate under controlled conditions.