Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DiffusionRIR: Room Impulse Response Interpolation using Diffusion Models

Sagi Della Torre, Mirco Pezzoli, Fabio Antonacci, Sharon Gannot

TL;DR

The paper addresses the challenge of reconstructing high-resolution Room Impulse Responses (RIRs) from sparse measurements by casting the RIR set as a 2D image and applying a diffusion-model inpainting approach. It introduces a RePaint-style DDPM framework trained on a small RIR dataset and demonstrates reconstruction of missing RIRs for linear and curved microphone arrays, outperforming spline cubic interpolation in NMSE and Cosine Distance. The method maintains robustness to large gaps and shows promising generalization across array geometries and room configurations, with analyses of EDC-derived $T_{60}$ and source-angle effects. This diffusion-based approach offers a data-efficient pathway to high-resolution acoustic mapping and can enable data augmentation and improved sound-field tasks with limited measurements.

Abstract

Room Impulse Responses (RIRs) characterize acoustic environments and are crucial in multiple audio signal processing tasks. High-quality RIR estimates drive applications such as virtual microphones, sound source localization, augmented reality, and data augmentation. However, obtaining RIR measurements with high spatial resolution is resource-intensive, making it impractical for large spaces or when dense sampling is required. This research addresses the challenge of estimating RIRs at unmeasured locations within a room using Denoising Diffusion Probabilistic Models (DDPM). Our method leverages the analogy between RIR matrices and image inpainting, transforming RIR data into a format suitable for diffusion-based reconstruction. Using simulated RIR data based on the image method, we demonstrate our approach's effectiveness on microphone arrays of different curvatures, from linear to semi-circular. Our method successfully reconstructs missing RIRs, even in large gaps between microphones. Under these conditions, it achieves accurate reconstruction, significantly outperforming baseline Spline Cubic Interpolation in terms of Normalized Mean Square Error and Cosine Distance between actual and interpolated RIRs. This research highlights the potential of using generative models for effective RIR interpolation, paving the way for generating additional data from limited real-world measurements.

DiffusionRIR: Room Impulse Response Interpolation using Diffusion Models

TL;DR

The paper addresses the challenge of reconstructing high-resolution Room Impulse Responses (RIRs) from sparse measurements by casting the RIR set as a 2D image and applying a diffusion-model inpainting approach. It introduces a RePaint-style DDPM framework trained on a small RIR dataset and demonstrates reconstruction of missing RIRs for linear and curved microphone arrays, outperforming spline cubic interpolation in NMSE and Cosine Distance. The method maintains robustness to large gaps and shows promising generalization across array geometries and room configurations, with analyses of EDC-derived and source-angle effects. This diffusion-based approach offers a data-efficient pathway to high-resolution acoustic mapping and can enable data augmentation and improved sound-field tasks with limited measurements.

Abstract

Room Impulse Responses (RIRs) characterize acoustic environments and are crucial in multiple audio signal processing tasks. High-quality RIR estimates drive applications such as virtual microphones, sound source localization, augmented reality, and data augmentation. However, obtaining RIR measurements with high spatial resolution is resource-intensive, making it impractical for large spaces or when dense sampling is required. This research addresses the challenge of estimating RIRs at unmeasured locations within a room using Denoising Diffusion Probabilistic Models (DDPM). Our method leverages the analogy between RIR matrices and image inpainting, transforming RIR data into a format suitable for diffusion-based reconstruction. Using simulated RIR data based on the image method, we demonstrate our approach's effectiveness on microphone arrays of different curvatures, from linear to semi-circular. Our method successfully reconstructs missing RIRs, even in large gaps between microphones. Under these conditions, it achieves accurate reconstruction, significantly outperforming baseline Spline Cubic Interpolation in terms of Normalized Mean Square Error and Cosine Distance between actual and interpolated RIRs. This research highlights the potential of using generative models for effective RIR interpolation, paving the way for generating additional data from limited real-world measurements.
Paper Structure (10 sections, 2 equations, 16 figures)

This paper contains 10 sections, 2 equations, 16 figures.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Proposed Approach
  4. Inpainting with Diffusion Models
  5. Image Representation of the RIR Set
  6. Experiments
  7. Experiment Setup
  8. Performance Measures and a Baseline Method
  9. Results
  10. Discussion

Figures (16)

  • Figure 1: Geometric setup of a room with a source and a microphone array. Measured and missing microphones are marked in green and red dots, respectively. We aim to reconstruct the RIR of the missing microphones.
  • Figure 2: Heatmap of RIR matrix $\mathbf{H}$, with one zoomed-in RIR view. Each column in the matrix represents one microphone.
  • Figure 3: NMSE for inpainted and baseline SCI vs. mask ratio.
  • Figure 4: CD for inpainted and baseline SCI vs. mask ratio.
  • Figure 5: Linear array configuration: Measured (blue) and missing (red) microphones.
  • ...and 11 more figures