Models for Real-Time Binaural Rendering using Church Impulse Responses

Abstract

Reproducing the spatial acoustics of a room is key to the feeling of spatiality of audio in headphones, with applications in, for example, cultural preservation. The propagation of sound between a point source and a receiver in a room is commonly modelled as a linear time-invariant system, where its room impulse response fully models how sound propagates. By combining the room impulse response and head-related impulse response, which model how a person's body filters sound, we can create a binaural room impulse response, which enables spatial audio reproduction via headphones. However, the room impulse responses are often long, leading to substantial computational demand for naive real-time processing.

This thesis looks into different modelling approaches for real-time filtering of room impulse responses measured in the Santi Marcellino and Pietro church, Cremona, Italy. Different models are used to investigate how room impulse responses can be represented compactly and computationally efficiently, while preserving accuracy for use in real-time processing. These methods include linear convolution, partitioned frequency-domain convolution, and a model exploiting the low-rank structure of the binaural room impulse responses.

Each model represents the impulse response differently and therefore gives rise to different trade-offs. Linear convolution preserves the full time-domain representation and serves as a natural reference. Partitioned frequency-domain convolution uses a blockwise frequency-domain representation, which provides the lowest computational cost but introduces an inherent latency. The low-rank model instead enforces a reduced-dimensional approximation, which, in turn, lowers storage requirements while offering a balance between computational cost and model complexity. These findings give rise to a general guideline for which application each method is most suited for.

Popular Abstract

Can we recreate the sound of a church in real time?

This thesis investigates whether it is possible to recreate spatial audio inside a large church, giving the illusion of sound bouncing off the walls and into your ears. If successful, anyone with a pair of headphones and access to a modern phone or computer could experience the acoustics of the church, an auditory experience that would otherwise be beyond their reach. However, doing this convincingly and in real time turns out to be more challenging than one might think.

By recording audio inside a church, we can capture what makes that room sound unique, both the direct sound that reaches your ears immediately and the echoes that arrive later after reflecting off walls, ceilings and other surfaces. The main challenge is processing all of this quickly enough that it feels natural. In theory, this could make it possible to filter your voice during a phone call so that someone wearing headphones hears it as if you were both standing inside a church.

Three different methods were proposed and tested. The first was simple and straightforward to implement, but too slow to meet the demands of real-time audio. The second achieved the same result much faster, and proved quick enough for real-time processing. The third compressed the audio data very efficiently and greatly reduced storage requirements, but still fell short of real-time performance. This made the second method the clear winner, and the only one able to achieve the main goal of the project.

One challenge the study encountered was how difficult it is to measure how convincing an audio effect sounds without involving human listeners. A computer can compare signals mathematically, but that does not always reflect what our ears actually perceive. As a result, listening and comparing by ear remained an important part of the evaluation, a reminder that, at least for now, the human touch still matters.

To complement the project, a simple computer program was also created so that curious listeners can test it for themselves and hear how their favourite song would sound inside the Santi Marcellino and Pietro church in Cremona.

• author: Persson, Olle
• supervisor: David Sundström; Oskar Keding
• organization: Mathematical Statistics
• alternative title: Modeller för binaural rendering i realtid med hjälp av kyrkoimpulssvar
• course: FMSM01 20261
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMS-3554-2026
• ISSN: 1404-6342
• other publication id: 2026:E39
• language: English
• id: 9230014
• date added to LUP: 2026-06-11 13:10:24
• date last changed: 2026-06-11 13:10:24