Room Geometry Estimation from Higher-Order Ambisonics Signals using Convolutional Recurrent Neural Networks

Poschadel N, Hupke R, Preihs S, Peissig J. Room Geometry Estimation from Higher-Order Ambisonics Signals using Convolutional Recurrent Neural Networks. In: AES Convention 150. Audio Engineering Society; 2021. Paper 10482. Available from: https://aes.org/publications/elibrary-page/?id=21075

@inproceedings{Poschadel2021_21075,
  author = {Poschadel, Nils and Hupke, Robert and Preihs, Stephan and Peissig, Jürgen},
  title = {{Room Geometry Estimation from Higher-Order Ambisonics Signals using Convolutional Recurrent Neural Networks}},
  booktitle = {AES Convention 150},
  note = {Paper 10482},
  year = {2021},
  month = may,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=21075}
}

TY  - CPAPER
TI  - Room Geometry Estimation from Higher-Order Ambisonics Signals using Convolutional Recurrent Neural Networks
AU  - Poschadel, Nils
AU  - Hupke, Robert
AU  - Preihs, Stephan
AU  - Peissig, Jürgen
T2  - AES Convention 150
M1  - Paper 10482
PY  - 2021
DA  - 2021/05/06
UR  - https://aes.org/publications/elibrary-page/?id=21075
PB  - Audio Engineering Society
LA  - en
AB  - Knowledge of room geometry is a fundamental component for modeling acoustic environments. Since most common methods for room geometry estimation are based on prior knowledge, the generalization to unknown environments is somewhat limited. Deep learning based approaches have delivered promising results for the blind estimation of acoustic parameters considering mainly monaural signals. The purpose of this contribution is to investigate the effect of multichannel higher-order Ambisonics (HOA) signals on the performance of a convolutional recurrent neural network for blind room geometry estimation. Therefore a HOA-dataset of noisy speech signals in simulated rooms with realistic frequency-dependent reflection coefficients is introduced. Results show that for each additional Ambisonics order the estimation performance increases with the fourth-order model achieving a mean absolute error of 1.24 m averaged over all three room dimensions.
ER  -