We describe our submission to Task A of the 1st DAFx parameter estimation challenge. The task is to recover the six physical parameters of a simulated metal-plate reverberator -- its dimensions and material properties -- from a single impulse response (IR). We treat this as a black-box optimization: candidate parameter sets are fed to the simulator and scored by a loss against the target IR. The method has two stages. The first uses CMA-ES, an evolutionary optimizer, to recover five of the six parameters, comparing IRs under an amplitude-normalized loss. Amplitude normalization makes the search robust but discards the cue to the sixth parameter, the plate's surface density; a second stage therefore estimates it alone, with a ternary search on the un-normalized loss. As the choice of loss strongly affects the search, we select it beforehand, and analyze why compression in the common multi-scale spectral loss degrades recovery. Finally, we test our method on a validation set of 50 IRs, discuss a pathological failure mode, and ablate to justify having two different stages instead of a unified CMA-ES search.
Figures & tables
Configuration
Geomean NMSE
Conv. rate
L1 -STFT (base)
4.9\times10−14
1.00
+ multi
4.8\times10−14
1.00
+ SC
1.0\times10−13
0.98
+ SC + multi
3.2\times10−14
1.00
+ log (bounded)
8.9\times10−4
0.40
+ log
1.3\times10−1
0.00
Table 1: Component decomposition from L1 -STFT to the MSS+SC variant (components defined in text); the challenge MSS baseline is the log + multi row, and log (bounded) is the variant log(1+x) . Values are geometric-mean NMSE and convergence rate ( NMSE<0.02 ) over the 50 IRs.
Window (pts)
Geomean NMSE
Conv. rate
512
5.9\times10−12
0.92
1024
3.7\times10−12
0.88
2048
1.8\times10−13
0.96
4096
4.9\times10−14
1.00
8192
9.2\times10−14
0.96
Table 2: Effect of STFT window size on recovery, for the single-resolution uncompressed L1 -STFT loss. Recovery saturates by 4096 points (bold, our choice); every window still outperforms the compressed losses of Table 1 .
Figure 1: Per-IR NMSE (log–log) before and after Stage 2 μ -refinement, over 50 generated IRs. All points except one (with pathological xo ) fall well below y=x , with refined NMSE often on the order of 10−14 or lower.
Figure 2: Distribution of per-IR Stage 1 runtime over the 50 validation IRs (linear axis, broken to show the tail). Most IRs finish near the median (32.6 s), while a few hard IRs extend past 200 s and a lone outlier reaches ∼640 s, pulling the mean (76.9 s) above the median.
All 50 IRs
Outlier in Fig. 1
Search
NMSE median
Time (s) median / mean
NMSE / Time (s)
Proposed
1.3\times10−14
33.1 / 76.9
7.6\times10−2 / 205
One-stage
2.0\times10−14
36.7 / 167
1.4\times10−2 / 3604
Table 3: Our two-stage pipeline vs. One-stage joint CMA-ES on the 50 validation IRs. Our pipeline is faster and more accurate. Both fail on the pathological outlier of Fig. 1 , with comparable error, but the one-stage search uses far more time.
We present a simulation-trained, non-iterative estimator for Task A of the 1st DAFx Parameter Estimation Challenge. Each unnormalized plate-reverb impulse response is summarized by amplitude, spectral, and decay descriptors, and an ensemble of tree regressors estimates the six target parameters in one pass. Across two independent synthetic validation sets, the normalized models outperform the training-set mean and an earlier raw-regression baseline. On a shared set, the final ensemble also outperforms a single run of the official default PSO at substantially lower inference cost. Since the official labels are hidden, parameter accuracy is measured on simulator-matched data, and the released responses support only audio-side consistency checks. The estimator returns point estimates without uncertainty.
Minhui Lu, Joshua D. Reiss
Centre for Digital Music Queen Mary University of London London, United Kingdom
We address Task A of the 1st DAFx Parameter Estimation Challenge, which aims to retrieve the physical parameters of a plate model from an impulse response. To do so, we use the Simulation-Based Inference (SBI) framework, in which we train a neural network to estimate a density over plate parameters given an impulse response, using a dataset generated by the simulator. Inference for a new impulse response then requires only a forward pass through the network, without involving the simulator. For each test observation, we fine-tune a specific network: additional simulation rounds are performed by sampling parameters from the current estimated distribution, simulating the corresponding impulse responses, and fine-tuning to produce the specialized network.
Dylan Sechet, Marc Evrard, Matthieu Kowalski
Laboratoire Interdisciplinaire des Sciences du Numérique Université Paris-Saclay, Inria, CNRS, CentraleSupélec
Task B of the 1st DAFx Parameter Estimation Challenge requires estimating the frequencies, decay rates, gains, and number of modes in a dense plate-reverb impulse response. Weak and overlapping modes make sparse peak detection prone to severe undercounting. We train an ExtraTrees regressor on simulator-generated data to predict mode counts in four frequency bands. These counts define dense frequency grids, after which a differentiable all-pole resonator model refines decay and gain while keeping frequency fixed. On two separate synthetic validation sets, the system reduces a local challenge-style error by about 66% relative to the official default peak-picking baseline. The improvement is mainly associated with lower mode-count mismatch, while decay and gain remain the largest error sources. These findings support separating modal-density estimation from continuous parameter fitting.
Minhui Lu, Joshua D. Reiss
Centre for Digital Music Queen Mary University of London London, United Kingdom