Humans make use of spatial hearing continuously to make sense of the world around. For example, to rapidly localize events in the environment and to communicate in noisy situations. Specifically, spatial hearing helps you to focus on a voice of interest in the midst of background noise (e.g. a colleague’s voice in the midst of voices of other colleagues and computers whizzing). Thus, spatial hearing is crucial for humans and the inability to localize sounds hampers communication in everyday life. Yet, it is still unknown how the human brain computes the location of real-life sounds in real-world listening situations because prior research concentrated on localization of simple sounds (for example, pure tones) in strictly controlled listening situations and experiments.
Importantly, knowledge of these brain mechanisms is needed to help hearing impaired listeners. HI listeners (over 34 million EU citizens and 5% of the worldwide population) experience great difficulties with understanding speech in noise environments. These problems persist even with an assistive hearing device such as a cochlear implant. Reduced spatial hearing contributes to these problems as it makes it difficult to filter out the voice of interest based on location information. As a result of these persistent communication problems, hearing impaired listeners are more prone to social isolation, low academic achievements, and unemployment. Besides the high personal impact, this also has a high economic impact on society.
In this research project, I took a novel approach by bringing together multiple scientific disciplines to address this problem. That is, the objectives of this Marie Sklodowska-Curie Action (MSCA) were (1) to develop a neurobiological-inspired deep neural network (DNN) model of location encoding of real-life sounds in the human brain; (2) to validate deep neural networks as models of sound location encoding in the human brain using measurements of neural activity; and (3) to employ the DNNs to investigate the neural representation of sound location in cochlear implant users and to develop signal processing strategies for cochlear implants that optimize subsequent spatial processing in the brain.
One of the main outcomes of the Action is a neurobiological-inspired convolutional neural network model (Objective 1). Our results show that such a model can accurately predict sound localization in the horizontal plane and that network localization acuity resembles human localization acuity for frontal locations. Crucially, the research outcomes highlight the potential of neurobiological-inspired deep neural network models as an approach to modeling human (spatial) hearing. Future neuroscientific research and clinical research is expected to benefit from the developed models, for example to assess neuronal sound location processing and to optimize signal processing strategies for cochlear implants that maximize subsequent spatial processing in the brain.
