eess.ASSep 28, 2026

Simulation-Based Inference for Plate Reverb System Identification

Authors: Dylan Sechet, Marc Evrard, Matthieu Kowalski

Organizations: Laboratoire Interdisciplinaire des Sciences du Numérique Université Paris-Saclay, Inria, CNRS, CentraleSupélec

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

ParameterSymbolRangeUnit
Areal mass densityμ\mu[2.43, 106.15][2.43,\ 106.15]kg m−2\mathrm{kg}\text{\,}{\mathrm{m}}^{-2}
Bending-to-mass ratioD/μD/\mu[0.28, 201.2][0.28,\ 201.2]m4 s−2{\mathrm{m}}^{4}\text{\,}{\mathrm{s}}^{-2}
Tension-to-mass ratioT0/μT_{0}/\mu[9.4×10−5, 411.5][9.4{\times}10^{-5},\ 411.5]m2 s−2{\mathrm{m}}^{2}\text{\,}{\mathrm{s}}^{-2}
Plate length ( yy )LyL_{y}[1.1, 4.0][1.1,\ 4.0]m\mathrm{m}
Readout position ( xx )xox_{o}[0.51, 1.0][0.51,\ 1.0]frac. LxL_{x}
Readout position ( yy )yoy_{o}[0.51, 1.0][0.51,\ 1.0]frac. LyL_{y}
Table 1: Ranges of the six estimated parameters, used to normalize the NMSE ( 2 ). The invariant ranges ( μ\mu , D/μD/\mu , T0/μT_{0}/\mu ) are those induced by the raw-parameter ranges [ 3 ] ; Lx=L_{x}=1\text{,}\mathrm{m}$$ is fixed.
Figure 1: Summary of the neural architecture.
Figure 2: Architecture of one CNN branch.
Figure 3: Sequential SBI refinement.
  Per parameter
 Globalμ\muD/μD/\muT0/μT_{0}/\muLyL_{y}xox_{o}yoy_{o}
PSO baseline50.5626.169.728.4680.4395.1883.39
Offline2.172.180.100.235.503.231.75
Refined3.802.490.030.1411.454.933.74
Selected0.470.680.020.081.660.150.25
Oracle0.340.440.030.081.250.100.15
Table 2: Average per-parameter NMSE ( ×10−3\times 10^{-3} , lower is better) on the 5050 -IR test set. PSO is the challenge’s particle-swarm baseline.
Figure 4: Distribution of the per-parameter NMSE across the 5050 -IR test set.
Figure 5: TARP expected coverage of the offline posterior over 500 held-out plates.

Explore similar work

CardsList

  1. Simulation-Based Plate-Reverb Parameter Estimation from a Single Impulse Response

    Aug 1, 2026Minhui Lu, Joshua D. ReissReverberationImpulse Response

    Centre for Digital Music Queen Mary University of London London, United Kingdom

  2. Band-Count Dense Modal Estimation with Fixed-Frequency Differentiable Resonator Refinement

    Aug 1, 2026Minhui Lu, Joshua D. ReissReverberationDensity Ratio Estimation

    Centre for Digital Music Queen Mary University of London London, United Kingdom

  3. An Introduction to Bayesian and Frequentist Simulation-Based Inference with Machine Learning

    Jul 23, 2026Maximilian Dax, Theo Heimel, Gilles LouppeBayesian InferenceBayesian

    ELLIS Institute Tübingen, Max Planck Institute for Intelligent Systems, Tübingen AI Center, Tübingen, Germany · CP3, Université catholique de Louvain, Louvain-la-Neuve, Belgium · Montefiore Institute, University of Liège, Belgium