PhD thesis, Imperial College London (2021)

 

Recent advances in machine learning have instigated a renewed interest in using machine learning approaches to better understand human sensory processing. This line of research is particularly interesting for speech research since speech comprehension is uniquely human, which complicates obtaining detailed neural recordings. In this thesis, I explore how machine learning can be used to uncover new knowledge about the auditory system, with a focus on discovering robust auditory features. The resulting increased understanding of the noise robustness of human hearing may help to better assist those with hearing loss and improve Automatic Speech Recognition (ASR) systems. First, I show how computational neuroscience and machine learning can be combined to generate hypotheses about auditory features. I introduce a neural feature detection model with a modest number of parameters that is compatible with auditory physiology. By testing feature detector variants in a speech classification task, I confirm the importance of both well-studied and lesser-known auditory features. Second, I investigate whether ASR software is a good candidate model of the human auditory system. By comparing several state-of-the-art ASR systems to the results from humans on a range of psychometric experiments, I show that these ASR systems diverge markedly from humans in at least some psychometric tests. This implies that none of these systems act as a strong proxy for human speech recognition, although some may be useful when asking more narrowly defined questions. For neuroscientists, this thesis exemplifies how machine learning can be used to generate new hypotheses about human hearing, while also highlighting the caveats of investigating systems that may work fundamentally differently from the human brain. For machine learning engineers, I point to tangible directions for improving ASR systems. To motivate the continued cross-fertilization between these fields, a toolbox that allows researchers to assess new ASR systems has been released.