Source-linked AI summary
A multi-resolution spectrogram approach for estimating the physical parameters of a plate reverb
Jared Lipkin, Meiying Chen, Benjamin R. Thompson, Andrea Cogliati, David A. Anderson, Michael C. Heilemann
TL;DR
The paper addresses estimating plate-reverb physical parameters from recorded impulse responses. It adapts ResNet-18 to multi-resolution spectrogram inputs with phase information and downsampling; the reported comparisons favor downsampled input, with phase further improving output-location estimates.
Problem
The paper addresses identifying derived physical parameters of a plate reverb from a recorded impulse response.
Source: §1 · p. 1Method
A modified ResNet-18 estimates plate parameters from multi-resolution spectrograms, with a phase channel and downsampling to focus on isolated resonances.
Source: §1 · p. 1, §2 · p. 1, §2.1 · p. 2Results
Downsampled input outperformed full-spectrum input, phase further improved output-location estimates, and the models had similar inference times.
Source: §3 · pp. 2–3Takeaways & Limitations
The authors conclude that image-processing machine-learning models can support audio system identification from spectral images, with phase and restricted bandwidth refining the inputs.
Source: §4 · p. 3Inference-time measurements on the 200-IR dataset may be affected by background CPU activity.
Source: §3 · p. 3
Abstract
from arXiv · showhide
The ResNet-18 image classification model is employed to determine the physical parameters of a plate reverb from a recording of the impulse response. The model is adapted to derive parameters using normalized and downsampled multi-resolution spectrograms computed from the provided impulse responses (IRs). To refine the prediction of the output location, the spectral phase response is also included as an additional input channel to the network since multiple output locations can give the same magnitude response for high-order resonant modes. On a 5000 IR validation set, our model achieves an average normalized mean squared error (NMSE) of 0.02920 across all parameters, with the lowest average NMSE occurring for parameters yo (0.00228), Ly (0.00347), and xo (0.00574).
1. INTRODUCTION
Task A estimates derived plate-reverb parameters from a recorded impulse response, using a modified ResNet-18 that processes audio transformed into spectrograms.
1. INTRODUCTION
- Task A identifies six derived plate parameters from a recorded impulse response, based on physical plate properties and specified constants.The stated physical properties include density, thickness, Young’s modulus, and tension.Source: §1 · p. 1
- The identified parameters include tension ratio T0/µ, plate height Lx, and output location (xo, yo).Source: §1 · p. 1
- A modified ResNet-18 processes audio transformed into two-dimensional spectrograms for parameter estimation.Source: §1 · p. 1
2. TASK-SPECIFIC MODIFICATIONS
The model adapts ResNet-18’s output and spectrogram inputs for plate-parameter estimation, including phase information to distinguish locations with matching magnitude responses. Preprocessing focuses on isolated resonances and uses downsampling to retain relevant modes at finer frequency resolution.
2. TASK-SPECIFIC MODIFICATIONS
- The submitted model’s output layer targets seven raw parameters, and default starting weights are bypassed.Source: §2 · p. 1
- Three channels encode spectrograms at N = 256, 512, and 1024, while a fourth carries N = 512 phase information to distinguish locations with matching magnitudes.For the (2,1) mode, the normalized locations (0.25, 0.5) and (0.75, 0.5) can share a magnitude response but differ in phase relative to other modes.Source: §2 · p. 1
2.1. Input Pre-Processing
- The model focuses on isolated resonances below the transition frequency, where modes are spatially well-defined and degeneracy is limited.Source: §2.1 · p. 2
- A 30× downsample to fs = 1470 Hz retains the first 20 resonances and yields 1.4355 Hz resolution, versus 43.0664 Hz before downsampling.The full-frequency spectrogram uses N = 512, 2048, and 8192 to align frequency resolution with the downsampled versions.Source: §2.1 · pp. 1–2
3. RESULTS
The submitted downsampled, phase-informed model performed best overall, with phase further improving output-location estimates. Downsampling outperformed full-bandwidth input, while inference times were similar across models on the 200-IR validation set.
3. RESULTS
- Full-bandwidth M1 beat the baseline by a large margin, but underperformed the downsampled models when NMSE, training cycles, and training time were considered.The authors suggest that poor modal separation at higher frequencies may leave little useful information about plate parameters.Source: §3 · p. 2
- The downsampled M1-Ds model performed notably better than full-spectrum M1, but marginally worse than the downsampled model with phase information.The authors attribute the advantage of downsampling to better frequency resolution where plate characteristics are most informative.Source: §3 · p. 2
- M1-DsPh was the submitted best-performing model, and its phase channel further improved output-location estimates over models without phase data.The output-location comparison is shown in Table 2.Source: §3 · pp. 2–3, Table 2 · p. 3
- Inference times were very similar across models on the 200-IR validation set, so M1-DsPh's increased accuracy came without a notable inference-time cost.The results were computed with eval.py on 200 five-second random IRs; Table 3 reports inference time.Source: §3 · pp. 2–3, Table 3 · p. 3
- The reference-dataset inference timings may be influenced by background CPU activity, so the authors consider the 200-IR inference time a more accurate representation of system performance.The reference data set is very small.Source: §3 · p. 3
4. CONCLUSION
The work demonstrates that image-processing machine-learning models can be used for audio system identification. It reports high accuracy for physical parameter estimation from spectral-image inputs and identifies phase information and low-modal-overlap bandwidth as refinements.
4. CONCLUSION
- The results demonstrate the viability of image-processing machine-learning models for audio system identification.The authors hope to extend the framework to empirical system identification on real plate-reverb units.Source: §4 · p. 3
- Converting audio signals to spectral images provides input features for high-accuracy physical parameter estimation.Source: §4 · p. 3
- The authors identify phase information and restricting frequency bandwidth to low-modal-overlap regions as ways to refine ResNet-18 input features.Source: §4 · p. 3