Algorithms for Sonifying Objectsinto Spatial Audio using Head-Related Transfer Functions

Sammanfattning

This thesis investigates algorithms for sonifying objects into spatial audio. The primary focus has been to examine how Head-Related Transfer Functions (HRTFs) could be interpolated while maintaining accurate spatial localization cues. More specifically, how different interpolation methods compare when interpolating spatial localization cues. The findings from the experiments showed that using a neural network to predict the spectrum of the HRTFs provide high spectral accuracy and is superior in capturing the localization cues. Furthermore, because human hearing cues are highly individual and depend on anatomical features such as ear and head shape, this thesis explores how a personalized neural network can improve HRTF interpolation. Utilizing the HUTUBS database, a model was trained to predict the individual spectrum of the HRTF. The results of the thesis showed that the model improved the shape of the magnitude interpolation. However, th etiming differences between filters were inaccurate. This thesis also investigates how the input layer of the anatomical features could be optimized to both increase prediction accuracy, but also to ease measurements for data collection. Overall, the findings demonstrate the potential of the proposed personalized model. However, further work is required to improve phase and timing prediction, as well as to expand the training dataset.