HRTF Generation for Data Demanding Machine Learning Algorithms

This thesis investigates the application of Machine Learning (ML) techniques to binaural audio research. Whilst there is plenty of work done in this domain currently, much of it is limited by the amount of available Head Related Transfer Function (HRTF) data required to train modern neural network-based ML models, resulting in researchers using a less data-driven approach or finding some workaround with the limited data. This thesis focuses on the generation of enough data to unleash the power of a wide variety of modern ML algorithms. A novel method is presented that can simulate unlimited realistic HRTFs using heads generated from Three-dimensional Morphable Models (3DMMs). The result has led to the creation of the HUman Morphable Model- based Numerically Generated Binaural Impulse Response Database (HUMMNGBIRD) database, created with the first 5000 HRTF sets generated by this method. Principle Component Analysis (PCA) and Variational Auto-Encoder (VAE) reconstruction models were created to investigate the potential of such a large amount of data. The results provide valuable insights into the research directions that could make good use of these types of artificially generated databases in the near future.