Communication Dans Un Congrès Année : 2023

From Room Impulse Responses to Wall Impulse Responses using Physics- Aware Deep Learning

Résumé

The problem of estimating the reflectivity of surfaces in a room, i.e., the wall impulse responses, using a set of room impulse responses is studied. The geometry of the room and setup and the device responses are assumed approximately known. It is shown that a purely learning-based method trained on simulated data performs well on simple matched conditions but shows poor generalizability to out-of-training scenarios. On the other hand, a purely model-based linear inverse method performs well in realistic conditions when the geometry is exactly known but is very sensitive to modeling errors and noise. A new physics-aware deep neural network architecture that takes the best of both solutions is hence proposed. The problem of correcting errors in the times of arrival of acoustic reflections is solved using an attention-based neural network. The linear inverse problem of estimating the wall impulse responses given the corrected model is solved by the conjugate gradient method, whose iterations are unrolled into a deep network. The two networks are interleaved and optimized end-to-end on a realistic simulated training set. This hybrid approach is favorably compared to the two baselines on out-of-training simulated data.

Fichier non déposé

Dates et versions

hal-04939466 , version 1 (10-02-2025)

Identifiants

  • HAL Id : hal-04939466 , version 1

Citer

Stéphane Dilungana, Antoine Deleforge, Cédric Foy, Sylvain Faisan. From Room Impulse Responses to Wall Impulse Responses using Physics- Aware Deep Learning. Proc. Forum Acusticum, 2023, Torino, Italy. ⟨hal-04939466⟩
112 Consultations
0 Téléchargements

Partager

  • More