Abstract
The objective of machine-based binaural sound source localisation is to estimate the position of sound sources as accurately as possible from a binaural sound signal. A subset of machine-based binaural sound source localisation methods utilise the listener’s measured head related transfer function (HRTF) dataset. These HRTF-based binaural sound source localisation methods provide opportunities for audio augmented reality and may also provide head mounted visualisations for the hard of hearing. In their original publications, the majority of HRTF-based binaural sound source localisation methods have been tested by synthetically generating their test data and training data using the exact same HRTF dataset. It has been shown that these methods are using the unique measurement noise present in the HRTFs in the dataset as a cue for localisation. This would not occur in a real life setting. Therefore, a comprehensive evaluation methodology is used in this thesis. The evaluation methodology is designed to test the methods’ robustness to different sources of noise associated with HRTF-based binaural sound source localisation.
A HRTF-based binaural sound source localisation method is proposed that uses interaural and spectral features for localisation of sound sources with any direction of arrival (DOA) on the full-sphere. The method is designed to be robust to the presence of reverberation, additive noise and different types of sounds. The method uses the interaural phase difference (IPD) for lateral angle localisation, then interaural and spectral features for polar angle localisation. For polar angle estimation, the method applies a different weighting to the likelihood distributions derived from the interaural and spectral features depending on the estimated lateral angle. In particular, only the spectral features are used for sound sources near or on the median plane. The proposed method outperforms the state of the art HRTF-based binaural sound source localisation methods in challenging testing conditions with the presence of measurement noise and reverberation.