Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RIR-Former: Coordinate-Guided Transformer for Continuous Reconstruction of Room Impulse Responses

Shaoheng Xu, Chunyi Sun, Jihui, Zhang, Prasanga N. Samarasinghe, Thushara D. Abhayapala

TL;DR

This paper tackles the challenge of reconstructing room impulse responses (RIRs) at unseen locations from sparse measurements. It introduces RIR-Former, a grid-free transformer that uses a sinusoidal position encoding for microphone geometry and a segmented multi-branch decoder to separately model early reflections and late reverberation, enabling one-step, forward inference. The key contributions are a geometry-aware, grid-free transformer architecture, a segment-aware decoder that balances different RIR components, and a fast, generalizable framework that outperforms PINN-RIR, DiffusionRIR, and SCI across varying missing rates in simulated environments. Ablation studies confirm the importance of the sinusoidal encoding and the segmented decoder. The results suggest strong potential for practical deployment with arbitrary array layouts and pave the way for extensions to complex geometries, dynamic scenes, and real-world validation.

Abstract

Room impulse responses (RIRs) are essential for many acoustic signal processing tasks, yet measuring them densely across space is often impractical. In this work, we propose RIR-Former, a grid-free, one-step feed-forward model for RIR reconstruction. By introducing a sinusoidal encoding module into a transformer backbone, our method effectively incorporates microphone position information, enabling interpolation at arbitrary array locations. Furthermore, a segmented multi-branch decoder is designed to separately handle early reflections and late reverberation, improving reconstruction across the entire RIR. Experiments on diverse simulated acoustic environments demonstrate that RIR-Former consistently outperforms state-of-the-art baselines in terms of normalized mean square error (NMSE) and cosine distance (CD), under varying missing rates and array configurations. These results highlight the potential of our approach for practical deployment and motivate future work on scaling from randomly spaced linear arrays to complex array geometries, dynamic acoustic scenes, and real-world environments.

RIR-Former: Coordinate-Guided Transformer for Continuous Reconstruction of Room Impulse Responses

TL;DR

This paper tackles the challenge of reconstructing room impulse responses (RIRs) at unseen locations from sparse measurements. It introduces RIR-Former, a grid-free transformer that uses a sinusoidal position encoding for microphone geometry and a segmented multi-branch decoder to separately model early reflections and late reverberation, enabling one-step, forward inference. The key contributions are a geometry-aware, grid-free transformer architecture, a segment-aware decoder that balances different RIR components, and a fast, generalizable framework that outperforms PINN-RIR, DiffusionRIR, and SCI across varying missing rates in simulated environments. Ablation studies confirm the importance of the sinusoidal encoding and the segmented decoder. The results suggest strong potential for practical deployment with arbitrary array layouts and pave the way for extensions to complex geometries, dynamic scenes, and real-world validation.

Abstract

Room impulse responses (RIRs) are essential for many acoustic signal processing tasks, yet measuring them densely across space is often impractical. In this work, we propose RIR-Former, a grid-free, one-step feed-forward model for RIR reconstruction. By introducing a sinusoidal encoding module into a transformer backbone, our method effectively incorporates microphone position information, enabling interpolation at arbitrary array locations. Furthermore, a segmented multi-branch decoder is designed to separately handle early reflections and late reverberation, improving reconstruction across the entire RIR. Experiments on diverse simulated acoustic environments demonstrate that RIR-Former consistently outperforms state-of-the-art baselines in terms of normalized mean square error (NMSE) and cosine distance (CD), under varying missing rates and array configurations. These results highlight the potential of our approach for practical deployment and motivate future work on scaling from randomly spaced linear arrays to complex array geometries, dynamic acoustic scenes, and real-world environments.
Paper Structure (9 sections, 8 equations, 5 figures, 3 tables)

This paper contains 9 sections, 8 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Proposed Method
  4. Experiments
  5. Experiment Setup
  6. Comparison Methods
  7. Evaluation Metrics
  8. Results
  9. Conclusion

Figures (5)

  • Figure 1: RIR reconstruction setup.
  • Figure 2: Known RIRs and their position embeddings are passed through an MLP for feature projection. The projected features are processed by a Transformer encoder, which captures spatial and contextual dependencies using self-attention mechanisms. The decoder, conditioned on the target position embedding $x$ and learned feature map $c$, consists of multiple MLPs that handle different segments of the RIR vector, followed by a final MLP that merges them to predict the unknown RIR at the desired location.
  • Figure 3: RIR reconstruction results, segmented into 8 parts and normalized within each segment for better visualization. The subfigures compare different methods: (a) SCI, (b) DiffusionRIR, (c) PINN, (d) Ours, and (e) Ground Truth.
  • Figure 4: NMSE and CD across different missing rates (Exp. 1 and 2).
  • Figure 5: Experiment setups. (a) Experiment 1: fixed source, fixed array center, and uniform point spacing. (b) Experiment 2: random source position, random array placement, and randomized point spacing within the ROI $\Omega$.