Hearing a room before it is built

Architects can show a client a building before it exists, down to the grain of the timber and the light at four in the afternoon. Until recently they could not let the client hear it with any confidence. That is changing on three fronts at once. Spatial audio has a complete international standard for listeners who move. Simulation that models sound as a wave, rather than as rays, has reached open datasets. And learned acoustic fields predict how a room sounds at places nobody measured. The question that remains is the one every simulation faces: how far beyond its measurements can the prediction be trusted?

A standard for sound you can walk through

In November 2025 ISO and IEC published ISO/IEC 23090-4:2025, the MPEG-I immersive audio standard. It specifies real-time interactive rendering of virtual and augmented reality audio for a listener who can move with six degrees of freedom, turning and walking through the scene, together with the metadata that describes the acoustic scene and a bitstream for storing and streaming it. The standard runs to 625 pages.

A standard matters because it fixes what a renderer must be told about a space for the sound to follow the listener. The room, its surfaces and its sources become data that can be carried from the tool that designs a space to the device that plays it.

Simulating sound as a wave

Most acoustic simulation has traced sound like light, as rays bouncing between surfaces. That is fast and works well at high frequencies, but sound bends around edges and interferes with itself, and at low frequencies those effects dominate. In October 2025 Treble Technologies released Treble10, more than 3,000 room impulse responses across ten furnished rooms, generated with a hybrid of wave-based and geometrical solvers. The authors' motivation is explicit: most simulated datasets rely on simplified geometrical acoustics and so fail to reproduce diffraction, scattering and interference. The dataset is public, so others can test against it.

Learning a room from a few measurements

The third front is learned acoustic fields: models that take a handful of measured responses in a room and predict the response anywhere else in it. NeRAF, presented at ICLR 2025, learns a radiance field and an acoustic field together, so that the 3D geometry and appearance recovered from images inform how sound is predicted at new positions, and reports state-of-the-art room impulse response synthesis with less data.

The GenDARA challenge at ICASSP 2025 put the idea to a practical test. Teams had to generate room impulse responses to supplement a small set of measured examples, and were ranked on how well those generated responses trained a downstream model that estimates how far away a speaker is. The winning team at MERL used a geometry-aware neural acoustic field, pre-trained on a large external dataset of room responses and then fine-tuned to each challenge room from its few measurements.

The comparison that decides

Across all three fronts, the direction is the same: a few real measurements, physics where it is reliable, and a learned model to fill the gaps. What decides whether the result can be trusted is the comparison it is scored against.

There are two honest baselines, and both are hard to beat. The first is the physics formula, calibrated to the room. The second is memory: the measured responses nearest to the point being predicted. Within a single well-measured room, memory is formidable, because the nearest measurement already carries the room's own quirks. A learned field earns its place where measurements run out: in an arrangement of the room nobody has measured, at the edge of the range, or in a room it has never seen.

Scores also need a scale a listener can use. A difference in reverberation time matters only if someone could hear it, so error is best expressed against the smallest difference a listener can detect. And predictions should be scored on positions or configurations held out before anyone looked, because a field fitted and checked on the same measurements has shown only that it can fit.

Why this matters

Rooms are expensive to change once built. A hall that sounds wrong, a classroom where speech blurs or an open office where every conversation carries costs years of complaints or a second construction budget. Prediction that holds where nothing was measured lets the decision be made while it is still a drawing. It also lets a venue test a performance before load-in, and lets sound follow a listener through a space that exists only as data.

Where Living Worlds fits

In Living Worlds, rooms, halls and valleys sound the way their stone, wood and air make them sound, zone by zone, and each acoustic model is checked against real measurements before it is relied on and used within the range where it held.

In a real variable-acoustics room, fitted on the middle 70% of configurations, Living Worlds beat memory alone on the 15% at each end, winning 6,846 of 7,869 held-out measurements (Met). In the public dEchorate room, it predicted each new arrangement before it was measured and beat memory of the nearest measurements in all 20 seeded orders, with errors near 10% of the decay time (Met, 20 of 20). In a second measured room neither registered condition was met: memory learned the room's own bias, and within a single room memory proved as good as learned practice (Not met).

In four public rooms never used before, measured at twelve positions each and scored against the smallest difference a listener can hear, the calibrated formula carried to new positions by a learned field beat the same formula corrected by its five nearest measurements at 1,089 of 1,282 positions and in all four rooms. The same study reports that the mean of the five nearest measured positions alone scored higher than the field in every room, so it does not show the field ahead of memory alone (Met, 4 of 4 rooms). Each study was registered before the data was opened and scored once.