Abstract
We present an algorithm that fully reverses the shoebox image source method (ISM), a popular and widely used room impulse response (RIR) simulator for cuboid rooms. Given a discrete multichannel RIR generated by the shoebox ISM for a microphone array of known geometry, the algorithm reliably recovers 18 input parameters: the 3D source position, the 3 room dimensions, the 6-degrees-of-freedom room translation and orientation, and an absorption coefficient for each of the 6 boundaries. The approach combines gridless image source localization with new procedures for room axes recovery and first-order-reflection identification. Extensive simulations demonstrate near-exact recovery for a 32-element spherical microphone array at a 16 kHz sampling rate, with estimation errors decaying as array size and sampling rate increase. This work represents the first algorithmic demonstration that this difficult inverse problem is fully solvable over a wide range of configurations.
Introduction
In acoustic signal processing, "hearing the shape of a room" is a classic inverse problem. While the forward problem—simulating sound behavior in a box-shaped room—was effectively solved by Allen and Berkley's Image Source Method (ISM) in 1979, the inverse problem remains notoriously difficult. Key challenges include overlapping echoes in band-limited signals, the difficulty of labeling reflections (distinguishing first-order from higher-order bounces), and the combinatorial explosion of possible echo-to-wall assignments.
This paper investigates whether the shoebox ISM model is algorithmically reversible. We demonstrate that given the output of the simulator, it is possible to reliably recover every single input parameter, providing the first empirical evidence that the ISM forward model is in-principle fully invertible.
Methodology
The proposed pipeline consists of three primary stages designed to overcome the limitations of traditional grid-based search methods.
1. Gridless Localization
To avoid the computational expense and "basis mismatch" of discrete 3D grids, the algorithm employs a super-resolution approach using the sliding Frank-Wolfe algorithm. This technique recovers the 3D coordinates of hundreds of "image sources" (virtual sources representing reflections) from the multichannel RIR with high precision.
2. Room Orientation Recovery
A critical insight of this work is that image sources in a cuboid room naturally form an orthogonal grid. The algorithm exploits this global structure to find the room's axes before labeling any echoes. A novel scoring function, $J_3$, rewards directions orthogonal to the vectors connecting pairs of image sources. By maximizing this score, the room's orientation is recovered from the point cloud of image sources.
3. Labeling and Parameter Inference
With the axes known, the algorithm identifies the true source (the one closest to the array) and casts "cones" along the recovered axes to isolate the nearest first-order reflections. These reflections are then used to calculate the room dimensions, translation, and the specific absorption coefficients for each of the six walls.
Experimental Evaluation
The algorithm was tested using 200 randomly generated cuboid rooms (ranging from $2\times2\times2$m to $10\times10\times5$m). Simulations utilized a 32-element spherical microphone array (modeled after the Eigenmike) and an 8-element non-spherical array for baseline comparison. Parameters including source/array placement, room rotation, and wall absorption (0.01 to 0.3) were fully randomized. RIRs were generated using a custom ISM implementation and the pyroomacoustics package.
Results and Discussion
The experiments reveal that the ISM is a mathematically structured transformation that preserves enough information to be fully undone, provided the microphone array has sufficient resolution.
- High Precision: At a 16 kHz sampling rate with a 32-element array, the algorithm achieved mean errors in room dimensions of only a few millimeters. Room orientation was recovered with an average error of less than $0.06^\circ$.
- Sampling and Scale Effects: As shown in the figure below, estimation errors for dimensions, source position, and orientation decay toward zero as the sampling frequency and array radius increase.
- Baseline Comparison: The method significantly outperformed the landmark Euclidean Distance Matrix (EDM) approach, correctly retrieving all first-order reflections in 95.5% of cases, compared to 25.5% for the baseline.
- Extrapolation: Once parameters are recovered, the algorithm can simulate RIRs for new source positions with high fidelity (Signal-to-Error Ratio > 20 dB).
Conclusion and Future Work
While currently limited to "vanilla" ISM scenarios (perfectly cuboid rooms and specular reflections), this work proves the algorithmic reversibility of the shoebox model. Future research will focus on extending the model to handle real-world complexities, such as non-convex geometries, frequency-dependent wall responses, and the presence of furniture or other diffractive objects.
Summary updated