Acoustic impulse response prediction using regularized regression models
Examensarbete för masterexamen
Different methods of calculating the impulse response of rooms have been developed over the years and is an important measure within e.g. room acoustics and building acoustics. In this thesis work, two different regularized regression models were evaluated to see if it is possible to estimate the rooms’ impulse response simply by knowing some features of the room. This way, the room impulse response would be obtained in a much more simple and straightforward way compared to the already existing methods. The two models that were evaluated in this thesis work are called LASSO (Least Absolute Shrinkage and Selection Operator) and ridge regression, these models both generate a weight vector, given a training set and a target. Considering a large data set of different room properties as the training set, the target describes a new set of room properties. The aim was to be able to describe the target as the superposition of the training set multiplied with the weight vector, generated by the models. This weight vector was then tested to see if the relationship also could be applied to the impulse responses, by the superposition of the impulse responses in the training set multiplied with the same weight vector. One difference between the two models is that the LASSO model can shrink the coefficients belonging to the less important features to zero, while for ridge regression, the coefficients can only get close to zero. This is what encourages sparsity in the LASSO model, which turned out to be a winning concept. Results showed that the LASSO model estimated room impulse responses around 20 dB better than the ridge regression model. The results also show that there is potential for these models with some adjustments within the model, but also by weighing the features in order of importance.
Room impulse response, LASSO, ridge regression, acoustics , regression models, MCRoomSim, pyroomacoustics, sparse representation