This project is our first attempt at a spatial extension to our modal-reverberation framework — a reverb in which the listener can move continuously through a captured real-world space in real time, at minimal CPU cost.

The venue

The Bunkervik — also known as “Il rifugio delle idee” — is a former wartime air-raid shelter in central Brescia, Italy. Built in the 1940s to shelter civilians from Allied bombing, it consists of a single, narrow, vaulted gallery roughly 30 m long, partially excavated below street level, with unfinished masonry and exposed concrete walls. Since 2016 it has been progressively reopened as a contemporary-art and performance space.

The venue came to our attention after sound artist Matteo Gualeni recorded a series of drum performances inside the tunnel and became interested in capturing its acoustic signature in a faithful but musically usable form.

What makes the Bunkervik an ideal testbed is its geometry: it is well approximated as one-dimensional. A listener walking along the central axis experiences a smooth, monotonic evolution of the reverberant signature as a function of a single longitudinal coordinate. That lets the whole “spatial” problem be parameterised by one scalar — the listener’s position along the tunnel.

The measurement campaign

The acoustic survey was carried out in November 2025 using the exponential sine-sweep method.

A single Genelec studio monitor served as the excitation source, placed at one end of the tunnel and rotated through six cardinal orientations (up, down, front, back, left, right) to reconstruct an approximately omnidirectional source by averaging. The receiver — a Beyerdynamic free-field omnidirectional condenser microphone — was placed at three positions along the central axis, at roughly 8 m, 10 m, and 25 m from the source. Audio was recorded at 48 kHz / 24-bit through a portable USB interface.

From impulse responses to modal data

Each measured impulse response was converted, offline in MATLAB, into a modal representation — the impulse response described as a sum of damped sinusoids, each with its own frequency, decay rate, and weight.

This representation has a useful property. The poles of the system — the frequencies and decay rates — are essentially a property of the room’s geometry and boundary losses. Only the weights (residues) depend on where the source and receiver sit. If the poles can be pinned down once, the position-dependent part of the problem collapses to just the weights.

The three impulse responses were each fitted independently, yielding between roughly 6,000 and 16,000 stable modes per position. The mode count and density grow with distance from the source, reflecting the increasingly diffuse character of the late reverberant tail at the far end of the tunnel.

A common modal basis

The catch: three independent fits produce three pole sets that are similar but not identical. The same physical mode gets identified at slightly different frequencies across positions, which makes interpolating the weights ill-defined — you can only interpolate two weights if they are attached to the same pole.

The solution is a common modal basis. The three pole sets are merged: poles are clustered along the frequency axis with a frequency-dependent tolerance, and each cluster is collapsed to a single representative pole. This produces one unified, position-invariant set of about 6,300 poles spanning 70 Hz to 12 kHz.

With that fixed basis in place, the per-position weights — plus a short FIR correction filter for each position — are re-fitted jointly. The result, for each measurement position, is a set of weights and filter taps attached to the same pole set. That shared structure is exactly what makes interpolation possible.

 

The plugin

For any listener position between the measured points, the weights and FIR taps are obtained by simple linear interpolation of the three measured sets — while the pole set stays fixed. Because the poles never change, the bank of resonators that does the actual audio work is precomputed once; moving the listener at run-time requires interpolating from arrays of 6 thousand elements. A propagation delay that tracks the source-to-listener distance is added on top.

The practical payoff: a moving listener costs almost nothing extra compared to a static modal reverb. By contrast, a convolution-based equivalent would have to crossfade between many static impulse responses. Preliminary benchmarks show a single instance running at around 25% of one CPU core on an Apple M2, with modulation active.

The whole architecture is exposed through a single-pane JUCE plugin, Bunkervik Spatial Reverb.

The interface is organised into five blocks that affect the sound:

• Aura — a high-frequency “modal warping” of the pole frequencies, giving tonal control in the brilliance band without altering modal density or the spatial structure.
• Spatial — the core position controls: a src → 30 m strip that moves the listener through the tunnel, a mic pos knob exposing the same parameter for DAW automation, a mic rot control that biases the interpolation between source orientations, and a global damping multiplier on all the decay rates.
• Mix — wet-bus gain and a dry/wet blend, defaulting to unity for use on a send.
• Modulation — two LFOs routed through a depth matrix. The factory routing slowly modulates mic position (a ~25 s cyclic walk up and down the tunnel) and aura boost (slow timbral breathing).
• Output — stereo metering and a sample-accurate bypass for A/B comparison.

The modulation block is what makes the spatial-modal architecture musically alive: slowly modulating the listener position drives the interpolation continuously and audibly translates the listener along the tunnel, while the precomputed pole structure keeps modal frequencies and decays perfectly stable. CPU cost is independent of modulation depth and rate.

Here’s a short video processing a drum loop:

 

What’s next

This case study was deliberately chosen to expose the mechanism in its simplest form — a single longitudinal coordinate, three measurement positions. Future work will extend the common-basis approach to two- and three-dimensional source–receiver grids, add source-position dependence, and run a formal perceptual evaluation against other spatial-reverberation methods.

---

Funded by the European Research Council (ERC) under EU Horizon 2020, Grant No. 950084 NEMUS. The authors thank the Comune di Brescia and Associazione Culturale Palazzo Monti for access to the Bunkervik venue.