The pinna is well known to provide humans with the ability to determine the elevation of sound sources by imposing direction-specific spectral changes to the sound. Recent Cardiff research (Stevenson-Hoare et al. 2022) showed that the pinna also enhances horizontal angular discrimination of sound direction in the frontal hemifield (but not the rear). The result is remarkable because localisation in the horizontal plane is supported by powerful binaural cues, particularly interaural time delays (ITDs), that were thought to overwhelmingly dominate performance, but we found a strong influence of the pinna during binaural listening, especially for oblique angles, indicating that binaural cues and pinna cues are being combined by the brain.
It is unclear what perceptual cues underlie this effect, i.e., what spectral features are responsible for it, and whether these spectral features are analysed monaurally or binaurally. In the latter case, their effects would be mixed in with the second binaural cue, interaural level differences. Since acoustic spectra are multidimensional (there are different levels for each frequency and ear), it is impossible to make reliable inferences from visual inspection or statistical analyses; there are spectral changes with source direction at both ears and for sources in both hemifields. It is also difficult to further explore the relevant cues psychophysically. The natural approach would be to isolate the spectral cues under study, but without the usual binaural cues, spatial perception will largely collapse.
The project will tackle this problem by first using machine learning to infer the relevant spectral cues. Neural networks will be trained to perform the sound localisation without using ITDs. They will therefore not be disturbed by measures to exclude ITDs during the testing phase. The network can then be tested to assess its ability to discriminate directions using one ear or both, and for sources in the front and rear hemifield. An important manipulation will be to explore the effects of changing the source spectrum, because if the source spectrum is uncertain, the second ear in a binaural presentation may be used as a reference.
Different network architectures will be tested, but it is expected that a straightforward classification network with a single hidden layer with a limited number of nodes should be sufficient. Limiting the size of the hidden layer will provide the opportunity to reverse engineer and infer spectral cues. Networks will be trained on sources across the horizontal plane with
1) fixed source spectra for both ears
2) fixed source spectra for one ear
3) variable source spectra for both ears
4) variable source spectra for both ears
Performance in cases (1) and (2) will be compared to see whether the spectra at the ear contralateral to the sound source is useful when the source spectrum is known. Performance in (3) and (4) can be compared to see whether the contralateral ear still provides a reference for the source spectrum when it is unknown. Using (1), the effect of the pinna on the spectra will be tested; the components of the spectra caused by headshadow and by the pinna can be separated acoustically and so tested separately (as in Stevenson-Hoare et al., 2022). The mapping from the input layer to the hidden nodes and the pattern of response of the hidden layer to different source directions will be examined for evidence of specific cue extraction. Networks trained with different numbers of hidden nodes will be explored to force the network to employ identifiable cues.
Identifying the cues that a neural network will learn from the available information does not prove that humans are adopting the same cues. To establish this, we will investigate human performance using virtual acoustic simulation of the minimum-audible-angle task used by Stevenson-Hoare et al. Stimuli presented to each ear over headphones will be spectrally filtered to simulate the cues discovered by the neural network and determine their influence on performance. To evaluate monaural processing the filtering will be applied only to the side ipsilateral with the sound source, and for binaural processing, both sides will be filtered.
Using these techniques, the project will establish how the brain exploits the effects of pinna filtering and integrates this information with coherent binaural cues to optimise sound localisation.
Home students are UK Nationals and EU students who can satisfy UK residency requirements (students must have been in the UK for >3 years before start of course).
As only a limited number of studentships are available across the Open School competition and a very high standard of applications is typically received, the successful applicants are likely to have a very good first degree (a First or Upper Second class BSc Honours or equivalent) and/or be distinguished by having relevant research experience.