Classification of room characteristics based on Room Transfer Functions (RTFs) relates to various topics in acoustics, e.g. room response equalisation and forensic audio analysis.
In this Master’s thesis, convolutional neural network models are proposed for room volume and wall admittance classification. The proposed models are trained using supervised learning based on simulated RTFs from a sound field model derived from the Helmholtz equation for wave propagation in small rooms. The classification performance of the models benefit from having microphones placed in structured grids
compared to randomly placed microphones. For classification of the room volume, a classification accuracy above approximately 90% for more than one microphone is achieved, while for wall admittance classification the accuracy is above 80%. The models for room volume classification were more sensitive to additive noise than the wall admittance models, which still provided better then initial accuracy at 0dB SNR.