Spherical Maps of Acoustic Properties as Feature Vectors in Machine-Learning-Based Estimation of Acoustic Parameters

Falcón Pérez R, Götz G, Pulkki V. Spherical Maps of Acoustic Properties as Feature Vectors in Machine-Learning-Based Estimation of Acoustic Parameters. J Audio Eng Soc. 2021;69(9):632-643. doi:10.17743/jaes.2021.0011

@article{FalconPerez2021_21460,
  author = {Falcón Pérez, Ricardo and Götz, Georg and Pulkki, Ville},
  title = {{Spherical Maps of Acoustic Properties as Feature Vectors in Machine-Learning-Based Estimation of Acoustic Parameters}},
  journal = {Journal of the Audio Engineering Society},
  volume = {69},
  number = {9},
  pages = {632--643},
  year = {2021},
  month = sep,
  publisher = {Audio Engineering Society},
  doi = {10.17743/jaes.2021.0011},
  url = {https://doi.org/10.17743/jaes.2021.0011}
}

TY  - JOUR
TI  - Spherical Maps of Acoustic Properties as Feature Vectors in Machine-Learning-Based Estimation of Acoustic Parameters
AU  - Falcón Pérez, Ricardo
AU  - Götz, Georg
AU  - Pulkki, Ville
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 69
IS  - 9
SP  - 632
EP  - 643
PY  - 2021
DA  - 2021/09/06
DO  - 10.17743/jaes.2021.0011
UR  - https://doi.org/10.17743/jaes.2021.0011
PB  - Audio Engineering Society
LA  - en
AB  - This work suggests a method of presenting information about the acoustical and geometric properties of a room as spherical images to a machine-learning algorithm to estimate acoustical parameters of the room. The approach has the advantage that the spatial distribution of the properties can be presented in a generic and potentially compact way to machine learning methods. The estimation of reverberation time T60 is used as a proof-of-concept study here. The distribution of absorptive material is presented as a spherical map of feature values in which each value is formed by calculating the equivalent absorption area visible through the corresponding facet of a polyhedron as seen from the polyhedron’s center point. The pixel values are then used as feature vectors and the real measured T60 values of corresponding rooms are used as target data. This work presents the method and trains a set of neural networks with different spherical map resolutions using a dataset composed of real-world acoustical measurements of a single room with 831 different configurations of furniture and absorptive materials. The estimation of reverberation time using the proposed approach exhibits a much higher accuracy compared to simple analytic methods, which proves the validity of the approach.
ER  -