Deep Room Impulse Response Completion -- A Multiexponential Decay Approach

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Authors

Date

2024-05-20

Department

Major/Subject

Acoustics and Audio Technology

Mcode

ELEC3030

Degree programme

CCIS - Master’s Programme in Computer, Communication and Information Sciences (TS2013)

Language

en

Pages

63+1

Series

Abstract

The acoustics of enclosed spaces, vital for applications like concert hall design, have been studied extensively since the 1960s. Emerging technologies such as augmented reality and virtual reality (AR/VR) have brought new demands on immersive audio, necessitating fast, realistic simulation of acoustics. Room Impulse Responses (RIRs), the acoustic fingerprint of a space, are used in the auralization of virtual spaces. However, generating accurate RIRs in real-time for dynamic AR/VR environments remains challenging. Traditional simulation methods such as geometrical acoustics and wave-based acoustics have known limitations in accuracy and computational cost. Hybrid methods combining geometrical acoustics and wave-based acoustics have been proposed, but they also face challenges, thus motivating the need for new RIR generation paradigms. This thesis introduces RIR completion, a novel RIR generation task to predict the remaining portion of an RIR from its early segment. A lightweight encoder-decoder style deep neural network, Deep Exponential Completion of Room impulse responses (DECOR), is proposed to efficiently complete RIRs. DECOR's highly informed decoder predicts the amplitude values of exponential decay curves within the multiexponential decay reverberation framework. Experimental results demonstrate DECOR's effectiveness and reliability compared to a state-of-the-art model, while achieving fewer audible artifacts and more than thirty-fold model size reduction. While this thesis focuses on deep learning-based implementation, it serves as an initial proof-of-concept for RIR completion, opening avenues for more refined future RIR completion methods.

Description

Supervisor

Välimäki, Vesa

Thesis advisor

Schlecht, Sebastian
Götz, Georg

Keywords

room impulse response, deep learning, room acoustics, room impulse response completion, multiexponential decay, generative acoustics

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202405263836

Collections

[dipl] Sähkötekniikan korkeakoulu / ELEC