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

Abstract

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.

1 Introduction

Modal representations are widely used in physical audio modeling and artificial reverberation because they express a resonant system as a sum of damped components. For a plate, these components are specified by modal frequencies, decay rates, and gains, or equivalently by a bank of resonant filters [1, 2]. The representation supports interpretable analysis and efficient resynthesis, but recovering it from a single dense impulse response is difficult. Modes can be closely spaced, weak resonances can be masked by stronger ones, and the mode count is unknown.

Task B of the 1st DAFx Parameter Estimation Challenge provides a benchmark for this problem. The two challenge tasks address different levels of the same generative process. Task A estimates a compact physical and observation description of the plate system, whereas Task B directly estimates the much larger modal representation whose components form the response. Each target response is generated by a public modal plate-reverb simulator available in the official challenge repository: The required output is a modal set

$$\mathcal{M}=\{(f_{m},\sigma_{m},g_{m})\}_{m=1}^{M},$$ (1)

where $f_{m}$ is modal frequency, $\sigma_{m}$ is the decay rate, $g_{m}$ is the modal gain, and $M$ is the identified number of modes. The evaluation penalizes frequency, decay, gain, and mode-count errors.

Existing estimators broadly follow either direct signal analysis or analysis-by-synthesis. Sinusoidal analysis estimates resonant components from the signal [3], while subspace and matrix-pencil methods estimate damped exponentials when the model order and signal conditions are suitable [4, 5]. Differentiable digital signal processing (DDSP) instead allows a structured signal model to be fitted by gradient descent [6, 7], but a large variable-size modal set still requires an initial structure. The official reference method uses spectral peak picking followed by local decay and gain estimates. It is inexpensive, but sparse peak detection can miss weak or overlapping modes. The specific challenge is therefore not only to calibrate individual modes, but also to construct a modal set with a plausible cardinality.

This paper asks whether band-wise mode-count estimation can provide the structural prior needed before local modal calibration. We propose a count-first system that predicts mode counts in broad frequency bands, places a dense modal grid, and applies bounded differentiable refinement to decay and gain. This decomposes variable-size estimation into learned count prediction, deterministic initialization, and constrained fitting. On two simulator-matched validation sets, the system substantially lowers local challenge-style error relative to the official default baseline. Most of the gain comes from reducing mode-count mismatch, while differentiable refinement adds a smaller improvement.

2 Method

Figure 1: Overview of Task B inference. Fixed response descriptors feed the trained band-count estimator, whose predictions define a dense modal initialization. A spectral objective through the differentiable resonator bank then updates bounded decay and gain correction factors while keeping $f_{m}$ fixed. Offline training of the count estimator is described in the text.

We use a hybrid computational inverse-estimation design because Task B requires both an unknown number of modes and continuous parameters for every mode. Supervised learning estimates band-wise mode counts, deterministic initialization places the frequencies and estimates decay and gain from the response, and constrained per-response optimization refines decay and gain. Offline, each simulated response $h_{n}^{\mathrm{syn}}$ is paired with its ground-truth modal list $\mathcal{M}_{n}^{\mathrm{gt}}$. A band-histogram operator $B$ converts the list to four count labels,

$$(h_{n}^{\mathrm{syn}},\mathcal{M}_{n}^{\mathrm{gt}})\mapsto(\phi(h_{n}^{\mathrm{syn}}),\mathbf{c}_{n}),\qquad\mathbf{c}_{n}=B(\mathcal{M}_{n}^{\mathrm{gt}}),$$ (2)

where $\phi$ is a deterministic 372-dimensional descriptor extractor with no learned parameters and is identical at training and inference. Robust-scaling statistics and four ExtraTrees regressors are fitted from these pairs. We denote the complete fitted predictor by $F_{\mathrm{ET}}:\phi(h)\mapsto\mathbf{c}$. At inference, $F_{\mathrm{ET}}$ and the deterministic initializer $I$ produce

$$\hat{\mathbf{c}}=F_{\mathrm{ET}}(\phi(h)),\hskip 18.49988ptI(h,\hat{\mathbf{c}})=\{f_{m}^{(0)},\sigma_{m}^{(0)},g_{m}^{(0)}\}_{m=1}^{\hat{M}},$$ (3)

where $\hat{M}$ is the sum of the bounded, rounded band counts. The final stage compares the target complex spectrum $H(\omega)$ with the resonator-bank spectrum $\hat{H}(\omega)$ and solves the constrained problem

$$(\hat{\boldsymbol{\alpha}},\hat{\boldsymbol{\beta}})=\arg\min_{\boldsymbol{\alpha},\boldsymbol{\beta}} \ \mathcal{L}\!\left(H,\hat{H}(\mathbf{f}^{(0)},\boldsymbol{\sigma}^{(0)}\!\odot\!\boldsymbol{\alpha},\mathbf{g}^{(0)}\!\odot\!\boldsymbol{\beta})\right)$$

subject to $$\alpha_{m},\beta_{m}\in[1/8,8],\hskip 18.49988ptf_{m}=f_{m}^{(0)},$$ (4)

where bold symbols collect the per-mode quantities and $\odot$ denotes element-wise multiplication. The method returns

$$\hat{\mathcal{M}}=\{(f_{m}^{(0)},\sigma_{m}^{(0)}\hat{\alpha}_{m},g_{m}^{(0)}\hat{\beta}_{m})\}_{m=1}^{\hat{M}}.$$ (5)

Only the target response is required at inference. The simulator and reference modal list are not used. Figure 1 summarizes the inference pipeline.

2.1 Band-wise mode counts

The frequency range is split into four bands: 20–200 Hz, 200–1000 Hz, 1–4 kHz, and 4–10 kHz. A four-output estimator predicts the number of modes in each band from 372 response descriptors. They summarize unnormalized amplitude, early/late energy, spectral shape, peak spacing, group delay, and band-limited decay, and are fixed rather than learned. After robust feature scaling, each output is predicted by an independent ExtraTrees regressor [8] with 500 trees and a minimum leaf size of 2.

Predicting mode counts per band lets the representation adapt its density to each target response rather than using a single fixed number of modes over the entire spectrum. This stage is the only supervised learning component in the Task B system.

2.2 Dense modal initialization

Given the predicted count in each band, modal frequencies are placed on a dense linear grid. Predicted counts are rounded, with at least eight modes per band and a 15,000-mode total cap. This is a count-conditioned approximation, not an analytical plate-mode enumeration: it uses neither the plate eigenfrequency equation nor explicit $(m,n)$ mode indices. Let $q$, in $\mathrm{dB\,s^{-1}}$, be the slope of a linear fit to the energy decay curve between $-5$ and $-35$ dB. With at least eight samples in this range, we set $\sigma_{\mathrm{base}}=\max(0.1,-q\log(10)/20)$. Otherwise, we use $\sigma_{\mathrm{base}}=2$. All decay rates are expressed in $\mathrm{s}^{-1}$. A mode in band $b$ is initialized as

$$\sigma_{m}^{(0)}=\sigma_{\mathrm{base}}s_{b}\left(1+\frac{0.5f_{m}^{(0)}}{c_{b}}\right),$$ (6)

where $\mathbf{s}=[20,20,20,40]$ contains dimensionless band scales and $c_{b}$ is the band center in Hz. The larger final-band scale and the factor 0.5 impose stronger high-frequency decay and a mild within-band increase, respectively. Initial gains are heuristically computed from the linearly interpolated complex DFT $H$ at each grid frequency,

$$g_{m}^{(0)}=-2\,\mathrm{Im}\{H(2\pi f_{m}^{(0)})\}\,\sigma_{m}^{(0)}T\sin(2\pi f_{m}^{(0)}T).$$ (7)

Here, $T$ is the sampling period.

2.3 Differentiable resonator refinement

The initializer is refined with a differentiable bank of second-order all-pole resonators. This follows the broader differentiable digital signal processing idea that a structured signal model can be optimized by gradient descent [6], while using a resonator form standard in modal sound synthesis and physical audio modeling [1, 2]. At angular frequency $\omega$, the modeled response is

$$\hat{H}(\omega)=\sum_{m}\frac{g_{m}z}{1-2r_{m}\cos(\omega_{m}T)z+r_{m}^{2}z^{2}},\qquad r_{m}=e^{-\sigma_{m}T},$$ (8)

where $z=e^{j\omega T}$ and $\omega_{m}=2\pi f_{m}$.

In the final system, the mode count and $f_{m}$ values are fixed by the dense initializer. Per-mode $\sigma_{m}$ and $g_{m}$ are updated through bounded multiplicative correction factors, so this stage acts as local calibration rather than full modal re-identification. For unconstrained optimizer variables $a_{m}$ and $b_{m}$, we use

$$\alpha_{m}=e^{\log(8)\tanh(a_{m})},\hskip 18.49988pt\beta_{m}=e^{\log(8)\tanh(b_{m})},$$ (9)

with $a_{m}=b_{m}=0$ initially. Thus both corrections start at one, remain within the eightfold trust region, and preserve the initial gain sign. The objective combines log-magnitude matching, a small phase-consistency term, and regularization:

$$\mathcal{L}= \ \underset{\omega}{\operatorname{mean}}\left|\log(|\hat{H}|+\epsilon)-\log(|H|+\epsilon)\right|$$

$$+\lambda_{\mathrm{ph}}\underset{\omega}{\operatorname{mean}}\left|u(\hat{H})-u(H)\right|$$

$$+\lambda_{\sigma}R_{\sigma}+\lambda_{g}R_{g},$$ (10)

where $\epsilon=10^{-12}$, $u(H)=H/(|H|+\epsilon)$ is the unit complex spectrum, $R_{\sigma}=\operatorname{mean}_{m}[(\log\alpha_{m})^{2}]$, and $R_{g}=\operatorname{mean}_{m}[(\log\beta_{m})^{2}]$. The final setting uses $\lambda_{\mathrm{ph}}=\lambda_{\sigma}=\lambda_{g}=0.02$. At each step, the resonator bank produces $\hat{H}$ on 1024 frequency samples from 20 Hz to 10 kHz. The loss is backpropagated through the bank to the decay and gain correction variables. frequency is not an optimization variable. Adam runs for 80 steps with learning rate 0.02 and gradient-norm clipping at 1, and the lowest-loss iterate is retained. The count estimator is implemented with scikit-learn, and the resonator refinement uses PyTorch automatic differentiation.

3 Experiments

Because the official modal lists are hidden, all development uses data generated by the public simulator. The count estimator uses 600 synthetic 5 s responses with a fixed 450/150 training/validation split. Modal-set accuracy is evaluated on two separately generated, disjoint 5 s sets: Validation 1 contains 8 responses and Validation 2 contains 12. Both are used to compare refinement bounds.

For modal-set experiments we report a local challenge-style score

$$\mathrm{RE}=\mathrm{RE0}+|M-\hat{M}|/M,\hskip 18.49988pt\mathrm{RE0}=(\mathrm{RE}_{f}+\mathrm{RE}_{\sigma}+\mathrm{RE}_{g})/3,$$ (11)

where $M$ and $\hat{M}$ are the reference and estimated mode counts. The first term averages the frequency, decay, and gain errors, while the second penalizes mode-count mismatch. Lower is better. Component-wise relative errors are clipped at one, and unmatched reference modes also contribute an error of one. To handle dense lists, modes are paired by greedy monotone frequency matching with a 0.5-octave threshold. This approximation is used only for local model selection. It is not the official hidden-label score.

The official default peak picker is the reference baseline. It uses 6 dB prominence, 2 Hz minimum spacing, and a 2 Hz prominence half-window over the same 20 Hz–10 kHz range. All modal-set variants use the same local matcher and score. For stage ablation, a total-count dense initializer predicts one mode count and allocates it to the four bands using fixed training-set fractions. We evaluate this initializer before and after fourfold decay/gain refinement, then replace it with the proposed direct band-count estimator and vary the refinement bound.

Runtime over the 16 released responses was measured on a MacBook Pro with an Apple M1 Pro chip (10-core CPU, 16 GB unified memory) running macOS 14.5, without GPU acceleration. Timing excludes synthetic-data generation and count-model training.

4 Results

Table 1 compares the complete system with the official default peak-picking baseline and separates the main system stages. Fourfold refinement improves the total-count dense initializer on both validation sets, and direct band-count prediction lowers RE further. Increasing the correction bound from fourfold to eightfold again improves both sets, whereas a sixteenfold bound does not transfer consistently. The eightfold setting is therefore retained. It reduces local RE by about 66% relative to the reference baseline on both validation sets.

Table 1: Local Task B relative error and stage ablation. Lower is better. Bold marks the column minimum.

Method	Validation 1	Validation 2
Official peak picking (default)	1.9699	1.9686
Total-count dense initialization	0.7364	0.7398
$+$ refinement (4$\times$)	0.6859	0.6935
Band-count $+$ refinement (4$\times$)	0.6738	0.6796
Band-count $+$ refinement (8$\times$)	0.6651	0.6750
Band-count $+$ refinement (16$\times$)	0.6627	0.6881

A major discrepancy in the reference baseline is mode count. It identifies an average of 69.0 and 67.3 modes on Validation 1 and 2, respectively, whereas the corresponding reference lists contain averages of 5240.1 and 6071.9 modes. The count-first system reduces the normalized count mismatch to about 5–6%.

On the separate 150-response count-validation split, ExtraTrees reduces both mean per-band and total-count MAE relative to Random Forest (Table 2), so it is retained in the complete system.

Table 2: Mode-count model selection on the 150-response count-validation split.

Model	Band MAE	Total MAE
Random forest	75.11	300.12
ExtraTrees	67.64	270.41

Table 3 decomposes the selected system’s remaining error. Normalized count mismatch is about 0.05, frequency error about 0.31, and decay and gain errors range from 0.69 to 0.85.

Table 3: Component errors for the selected eightfold setting.

Split	RE0	$\mathrm{RE}_{f}$	$\mathrm{RE}_{\sigma}$	$\mathrm{RE}_{g}$	$|\Delta M|/M$
Validation 1	0.6087	0.3176	0.6912	0.8172	0.0564
Validation 2	0.6215	0.3081	0.7015	0.8548	0.0536

For the final predictions on the 16 released responses, the mean refinement time was $6.51\text{\,}\mathrm{s}$ per response and end-to-end inference took $105.64\text{\,}\mathrm{s}$. The output contains approximately 83,800 estimated modes in total, with individual counts ranging from roughly 1,100 to 14,200.

5 Discussion

The comparison indicates that avoiding severe undercounting is central to the improvement over the official reference method. Peak picking returns only locally prominent resonances, whereas the plate simulator produces thousands of modes, including weak and overlapping components. Predicting counts in broad bands therefore provides a useful density prior before per-mode fitting. ExtraTrees improves the held-out count errors over Random Forest, although the remaining mismatch shows that count estimation is not solved completely.

Changing the decay/gain correction bound has a smaller effect than replacing sparse peak picking with count-first estimation. Increasing the bound from fourfold to eightfold improves both validation sets, while a sixteenfold range does not transfer consistently. This suggests that bounded calibration can correct part of the approximate, signal-informed initialization, but excess freedom can worsen the fit. Frequency error is unchanged because the grid is fixed, and decay and gain remain the dominant component errors.

Together, the findings support a hybrid division of roles: a learned estimator maps fixed descriptors to modal density, while a structured resonator model calibrates continuous parameters. This complements subspace methods that require a suitable model order [4, 5] and DDSP methods that optimize structured synthesizers [6, 7]. Here, the learned stage supplies the variable-size structure needed before differentiable fitting.

The evidence is limited to small validation sets generated by the same simulator family, and both sets were used when selecting the correction bound. The greedy monotone matcher is a memory-safe approximation rather than the official hidden-label evaluator, so the reported scores do not establish the challenge ranking. The official baseline comparison also does not cover tuned peak picking, subspace or matrix-pencil estimation, or unconstrained differentiable fitting. Finally, a misaligned grid frequency cannot be corrected by the present refinement. Future work should combine learned count priors with subspace-derived candidates or bounded frequency updates, improve decay/gain initialization, and estimate count uncertainty. Modal-list errors should also be compared with resynthesized response and perceptual measures, since these criteria need not improve proportionally.

6 Conclusion

This paper examined dense modal estimation from plate-reverb impulse responses. On local synthetic validation, predicting band-wise mode counts before fitting substantially reduces error relative to the official default peak-picking baseline, while bounded decay/gain refinement gives a smaller additional improvement. The results show that modal density is an important part of estimating large sets of weak and overlapping modes. Future work should allow limited frequency updates, improve decay and gain estimation, and compare with stronger classical modal-estimation methods and the official hidden-label evaluation.