Source-linked AI summary
Accurate Plate Reverb Parameter Estimation Using Two-Stage Evolutionary Search
Byunghoo Park, Jayeon Yi, Takyoung Kim, Minje Kim
TL;DR
The paper tackles recovery of six physical plate-reverberator parameters from a single simulated impulse response. It uses CMA-ES for five parameters with amplitude-normalized spectral loss, then ternary search for surface density using the unnormalized loss. On 50 validation impulse responses, the two-stage method reaches approximately 10^-14 NMSE in about 80 seconds per response, while log compression severely degrades recovery.
Problem
Task A asks whether six physical parameters of a simulated metal-plate reverberator can be recovered from one impulse response.
Method
The method combines CMA-ES recovery of five parameters under amplitude-normalized loss with ternary refinement of surface density using unnormalized loss.
Results
On 50 validation impulse responses, the method reaches NMSE values of approximately 10^-14 in about 80 seconds per response.
Takeaways & Limitations
Uncompressed L1-on-STFT-magnitude loss supports near-perfect recovery, whereas log-magnitude compression collapses recovery.
Takeaways & Limitations
A pathological case near xo = 1 causes simultaneous mode-amplitude errors and substantial surface-density drift, while direct six-dimensional search showed greatly reduced convergence.
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
The paper addresses Task A of the 1st DAFx challenge: recovering plate-reverberator physical parameters from a simulated impulse response. Because existing gradient-based or Newton-style approaches do not readily fit this rugged, large-scale problem, the authors use CMA-ES.
- Recovering simulator parameters can support learning useful configurations and broader creative applications of software plate reverberation.
- Task A requires inferring a plate’s physical parameters from a simulated impulse response, distinct from Task B’s modal estimation.
- Newton-based estimation has been demonstrated only on simple third-order systems and does not readily scale to this problem.
- Preliminary attempts to adapt gradient descent were unstable and prone to divergence, likely because the loss landscape is rugged and ill-conditioned.
- The authors therefore adopt CMA-ES, a derivative-free optimizer suited to the problem’s challenging objective landscape.
2. BACKGROUND
Task A inverts a plate simulator to estimate six derived physical parameters from an impulse response while holding other simulator parameters fixed. The simulator produces responses through parameter-dependent second-order IIR filters, making direct black-box parameter optimization appropriate.
- The plate simulator implements the general solution of the damped Kirchhoff–Love equation for a driven metal plate.
- Its physical parameter set includes material properties, plate dimensions, tension, decay constants, and input/output locations.
- The simulator output is a sum of second-order IIR filters whose coefficients depend on physical parameters, so the method optimizes parameters directly against the rendered response.
- Task A estimates {xo, yo, Ly, µ, T0/µ, D/µ}, where µ is surface density and D is flexural rigidity, while six other parameters remain fixed.
- The task design establishes a one-to-one correspondence between each impulse response and its six target parameters, enabling NMSE evaluation.
- CMA-ES samples candidate solutions from a multivariate Gaussian and adapts its mean, covariance, and global step size using evaluated candidates.
3. METHODOLOGY
The method uses two optimization stages: CMA-ES estimates five parameters under a normalized spectral loss, then ternary search estimates surface density from the unnormalized response. The design avoids direct six-dimensional search because preliminary experiments showed reduced convergence.
- 3. METHODOLOGY: The pipeline estimates six Task A parameters in two stages: CMA-ES first, followed by dedicated refinement of surface density µ.
- 3. METHODOLOGY: Both stages minimize uncompressed L1 distance between target and rendered STFT magnitudes using a 4096-point Hann window and 1024-point hop.
- 3.1. Stage 1: CMA-ES Search: Stage 1 searches seven raw parameters with multiple CMA-ES restarts, then derives and fixes xo, yo, Ly, T0/µ, and D/µ from a successful estimate.
- 3.1. Stage 1: CMA-ES Search: Peak normalization makes Stage 1 losses comparable across impulse responses and accelerates optimization, but removes the absolute-amplitude cue needed to estimate µ.
- 3.1. Stage 1: CMA-ES Search: Direct six-dimensional Task-A search was not pursued because preliminary experiments showed greatly reduced convergence, and complex raw-parameter-derived bounds require better out-of-bounds handling.
- 3.2. Stage 2: Ternary Search: Stage 2 performs 50 ternary-search iterations over µ ∈[2.43, 106.15] using the unnormalized loss while holding the other five estimates fixed.
4. LOSS ANALYSIS
The analysis identifies unbounded log-magnitude compression as the main cause of poor recovery relative to the uncompressed L1-on-STFT-magnitude loss. Multi-scale averaging and spectral convergence have smaller effects, while window size is secondary and bounded log variants remain inferior.
- The selected L1-on-STFT-magnitude loss outperforms the challenge’s multi-scale spectral loss baseline.The baseline combines log-magnitude L1 error across multiple STFT resolutions.
- Log compression alone eliminates convergence and raises geometric-mean NMSE from roughly 10^-14 to 10^-1.The multi-scale spectral baseline inherits this collapse, recovering only 2% of IRs.
- Adding spectral convergence to the baseline raises convergence from 0.02 to 0.68, but its geometric-mean NMSE remains 2.7 × 10^-4.This improvement still falls far short of the uncompressed L1-STFT loss.
- The degradation is specific to unbounded log(x): bounded log(1+x) and Smooth MSS are less damaging but still do not match the uncompressed loss.Their reported geometric means are 8.9 × 10^-4 and 7.9 × 10^-11, respectively.
- Window size has only a secondary effect: recovery is weakest at 512 points and saturates by 4096, while all windows outperform compressed losses.Even the weakest window setting, 512 points with geomean 5.9 × 10^-12, is an order of magnitude better than Smooth MSS.
5. METHOD VALIDATION AND SHORTCOMINGS
On 50 validation IRs, Stage 2 substantially reduced error at negligible additional cost, while a boundary-near pickup coordinate caused a pathological failure and one-stage CMA-ES was slower and slightly less accurate.
- Validation performance: 49 of 50 IRs improved after Stage 2, with median NMSE falling from 1.91 × 10−3 to 1.34 × 10−14.The refinement adjusted only µ; unnormalized L1-STFT loss also fell by four or five orders of magnitude.
- Runtime: Stage 1 required a median of 32.6 s and a mean of 76.9 s per IR, while Stage 2 averaged only 0.62 s.A few difficult IRs exceeded 200 s, and one reached approximately 640 s.
- Failure mode: The pathological outlier had ground-truth xo = 0.999980 near the boundary, where mode-amplitude sensitivity caused µ estimation to drift dramatically.The predicted and true values of 1 −xo differed by approximately one order, while the outlier's xo error was about 2 × 10−4.
- Ablation: The two-stage pipeline was faster and more accurate than one-stage joint CMA-ES overall, although both methods had comparable error on the pathological outlier.The one-stage method handled that outlier better only after exceeding an hour at the restart cap.
6. CONCLUSION
The paper presents a two-stage evolutionary method for recovering six plate-reverberator parameters from an impulse response. It combines normalized-loss CMA-ES for five parameters with unnormalized ternary refinement of surface density, achieving near-perfect validation recovery while exposing a boundary-related failure mode.
- Method: The method recovers five parameters with CMA-ES under amplitude normalization, then refines surface density µ using ternary search on the unnormalized loss.The loss is an uncompressed L1-on-STFT-magnitude objective.
- Loss selection: Uncompressed L1-on-STFT magnitude recovers parameters almost perfectly, whereas log-magnitude compression in the common multi-scale spectral loss collapses recovery.
- Results: On 50 validation IRs, the method reaches NMSE values of approximately 10−14 in about 80 seconds per IR.
- Limitations: The two-stage split is faster and more accurate than a single-stage joint CMA-ES search, but pickup coordinates extremely close to the plate boundary can trigger pathological failure.