From Numbers to Perception, Energy Decay Curves Prediction

Abstract: Predicting Room Impulse Responses (RIRs) remains a challenge due to the high dimensionality of audio signals and the need for perceptual accuracy. This paper introduces a neural network framework that predicts multi-band Energy Decay Curves (EDCs) directly from room geometry and material properties. Unlike standard models, our framework employs a custom composite loss function that optimizes for both energy levels and decay slopes in the log-domain. This ensures the predicted curves adhere to physical decay principles while maintaining high sensitivity to reverberation time and early reflections. Results demonstrate that the model successfully approximates ground-truth acoustics with minimal error in T30 and clarity indices. The approach offers a computationally efficient alternative to traditional simulations, facilitating realistic audio rendering for interactive virtual environments.

• The paper introduces a compact MLP–1D-CNN framework that predicts 24 one-third-octave energy decay curves from room geometry and material properties, reducing model size by 90% to about 9 million parameters and enabling a reported 5× inference speedup.
• The model uses a decibel-domain loss with a stride-based slope penalty to promote accurate, physically consistent decay, achieving R² values of 0.93 for T20, 0.90 for T30, and 0.67 for C50 on 6,000 simulated shoebox rooms.
• The approach supports real-time, frequency-dependent auralization through stochastic RIR reconstruction, but requires validation with measured rooms, non-rectangular geometries, and listening tests before its perceptual claims can be confirmed.

Overview

This paper presents a neural framework for predicting multi-band Energy Decay Curves (EDCs) directly from room geometry and material properties, targeting efficient, perceptually accurate auralization for interactive virtual environments. Rather than synthesizing full Room Impulse Responses (RIRs) — a high-dimensional task prone to phase inconsistencies and non-physical artifacts — the authors model EDCs as a compact intermediate representation and reconstruct RIRs downstream via a stochastic sign-assignment method. The work extends the authors' earlier LSTM-based broadband approach (Muhammad et al., 29 Sep 2025) with three principal contributions: a shift to 24 one-third-octave bands spanning 100 Hz–20 kHz, a 90% reduction in model size (from ~90M to ~9M parameters), and a physically motivated composite loss operating in the decibel domain.

Methodology

The dataset comprises 6,000 shoebox rooms simulated in Pyroomacoustics [8461310], parameterized by dimensions ($3$–$6\,\text{m} \times 3$–$6\,\text{m} \times 2.5$–$4\,\text{m}$), source–receiver distances of 1–4 m, and frequency-dependent wall absorption coefficients between 0.14 and 0.65 across all bands. The resulting $T_{60}$ distribution is deliberately non-uniform, concentrating training on typical rooms while retaining exposure to highly reverberant cases. Target EDCs are computed via the reverse-time Schroeder integral; inputs are MinMax-normalized.

The architecture consists of an MLP encoder compressing a 16-dimensional feature vector into a latent space, followed by a 1D-CNN decoder with three convolutional layers. Linear interpolation upsamples latent features to the target sequence length, which — together with a Sigmoid output bounding energies to $(0,1)$ — replaces the recurrent decoder responsible for "staircase" artifacts in prior work.

Training uses a composite loss in the log domain:

$$\mathcal{L}_{t} = \text{MSE}(\hat{y}_{dB}, y_{dB}) + \alpha \cdot \text{MSE}(\Delta \hat{y}_{dB}, \Delta y_{dB})$$

where $\Delta y_{dB}[n] = y_{dB}[n+k] - y_{dB}[n]$ is a finite difference with stride $k=50$, and $\alpha = 0.2$. The first term enforces absolute level accuracy; the slope penalty constrains the local decay rate, suppressing non-monotonic fluctuations and encouraging adherence to physical energy-dissipation behavior rather than point-wise memorization.

RIR reconstruction differentiates the predicted EDC into a magnitude envelope and assigns polarity via the Random Sign-Sticky (RSS) method with stickiness $6\,\text{m} \times 3$0, preserving spectral balance and low-frequency coherence.

Results

Evaluation on held-out simulated rooms yields strong agreement with ground truth across all 24 bands:

Parameter	RMSE	MAE	$6\,\text{m} \times 3$1
EDT (s)	0.10	0.07	0.79
$6\,\text{m} \times 3$2 (s)	0.06	0.04	0.93
$6\,\text{m} \times 3$3 (s)	0.07	0.05	0.90
$6\,\text{m} \times 3$4 (dB)	0.47	0.35	0.67

The central claim is that $6\,\text{m} \times 3$5 errors remain within the 5% Just Noticeable Difference (JND) threshold for the majority of test cases, implying perceptual indistinguishability from reference simulations in terms of reverberance. Clarity prediction is comparatively weaker ($6\,\text{m} \times 3$6), consistent with $6\,\text{m} \times 3$7's sensitivity to early-reflection structure that an envelope-level representation only partially captures.

Relative to the prior LSTM model, the ConvNet shows a modest regression in EDT MAE (0.033 s → 0.07 s). The authors frame this as a deliberate trade-off: the current model predicts 24 independent frequency bands versus broadband output previously, enabling representation of frequency-selective absorption by materials such as carpets or acoustic panels. The 10× parameter reduction also yields a reported 5× inference speedup, supporting real-time acoustic adjustment in VR scenes.

Limitations and open questions

Several constraints bound these results. All training data are synthetic shoebox simulations; generalization to non-shoebox geometries, vaulted ceilings, coupled volumes, diffraction, and non-uniform scattering remains untested, though the authors identify fine-tuning on measured RIR corpora as future work. The RSS reconstruction produces stochastic-phase RIRs whose perceptual validity rests on earlier validation (Muhammad et al., 29 Sep 2025) rather than listening tests conducted here — no subjective evaluation appears in this paper despite the JND-based claim. Additionally, the EDT accuracy loss relative to the broadband predecessor raises the question of whether per-band modeling inherently sacrifices temporal precision or whether this gap can be closed architecturally. Whether the slope penalty's fixed stride of 50 samples generalizes across reverberation times is also not examined.

Conclusion

This paper demonstrates that a compact 1D-CNN decoder, trained with a decibel-domain composite loss incorporating a stride-based slope penalty, can predict physically consistent, frequency-dependent EDCs from room geometry at 9M parameters — one-tenth the size of the preceding recurrent model — while keeping $6\,\text{m} \times 3$8 errors within perceptual JND bounds for most cases. Code, weights, and dataset scripts are released publicly, supporting reproducibility. The main open issues are validation on real measured acoustics and non-rectangular geometries, and formal subjective confirmation of the perceptual claims.