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

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.

💡 Research Summary

Room impulse responses (RIRs) are fundamental descriptors of an acoustic environment, underpinning applications such as acoustic‑based room design metrics, sound‑source localization, and immersive audio for virtual/augmented reality. Obtaining densely sampled RIRs throughout a space, however, is prohibitively time‑consuming and labor‑intensive, especially in large or geometrically complex rooms. Traditional reconstruction techniques—kernel ridge regression, parametric sound‑field models, compressive‑sensing‑based sparsity methods—rely on explicit physical priors and often break down in highly reverberant or irregular spaces. More recent learning‑based approaches (GANs, CNNs, physics‑informed neural networks, diffusion models) improve performance but typically assume uniform linear arrays (ULAs) or fixed grids, require per‑scene fine‑tuning, and suffer from high inference latency due to iterative denoising steps. Moreover, many of these methods treat the RIR as an image, discarding the distinct statistical characteristics of early reflections versus late reverberation.

The paper introduces RIR‑Former, a grid‑free, one‑step feed‑forward architecture that directly incorporates microphone geometry into a transformer backbone and explicitly separates the temporal structure of RIRs via a segmented multi‑branch decoder. The key contributions are threefold:

1. Sinusoidal positional encoding of microphone coordinates – each 3‑D microphone location (x_m) is mapped to a high‑dimensional periodic token (\gamma(x_m)) using sine and cosine functions across multiple frequencies (i = 6). This transforms low‑dimensional geometry into a richer feature space, enabling the transformer to learn scale‑invariant spatial relationships and to generalize to unseen array configurations.

2. Transformer encoder with global self‑attention – the concatenated token (