Contribution
Data-Driven Plane Wave Decomposition on Early Room Impulse Response Patterns
* Presenting author
Abstract:
In our DAGA 2023 paper, direction-of-arrival and time-of-arrival of early room reflections were estimated from multichannel room impulse responses (RIRs) using neural networks (NNs). This can be referred to as data-driven plane wave decomposition. The NNs were trained with artificial RIRs generated by superimposed plane wave patterns. The NNs were then tested on RIRs of simple shoebox rooms synthesised by the image source method (ISM). To further examine the prediction performance, enhanced shoebox rooms (ESBRs) that account for a more reasonable representation of real room characteristics were synthesised within a master thesis. A dedicated NN, trained with artificial plane wave RIRs showed high prediction performance on unseen synthesised RIRs from these ESBRs. For synthesis of train and test data, ideal point source excitation was considered. Additionally, RIRs of a real shoebox room were captured and used as test data. The prediction on very early room reflections yielded reasonable results. However, a considerable degradation of performance was observed compared to the prediction on synthesised RIRs of ESBRs. The dodecahedron speaker, used for measurement, causes impulse response shapes that deviate from those of an ideal point source. Thus, the model did not learn the shapes that are present in the measured RIR.