Machine learning prediction of room absorption coefficients from room impulse responses

Participating speaker: Martin Karlsson (MWL, KTH).

Abstract: The determination of absorption coefficients aims to characterize the acoustic properties of a room. From the absorption coefficients of a room other acoustical parameters can be derived. These include reverberation time as well as room acoustic parameters such as centre time and clarity, which depend on both the room geometry and the absorption. These parameters can be used to assess the suitability of a room for its intended acoustic purpose. Absorption coefficients can be determined in several ways. Experimental methods are typically based on the reverberation time or the energy decay of the room. In this work, a machine learning approach is presented for estimating absorption coefficients in non-convex rooms for each octave band from 125 Hz to 4000 Hz. A multilayer perceptron (MLP) is used and trained on data from a pretrained model called 'Echoscan' (ES) that predicts room geometry based on the room's impulse response. The data consist of the predicted room geometry and feature data from the MA-unit. Also, the Echoscan model is fine-tuned to allow the MA unit to improve the prediction of absorption coefficients in the MLP. After training the models on 100,000 different simulated rooms, the MLP was found to be unable to accurately predict the room absorption coefficients from the provided feature representation, suggesting that the provided input data are insufficient for reliable absorption coefficient estimation.