Auditory models of modulation perception in the spectral and temporal domain

Hearing emerges from a series of non-linear transformations that occur throughout the auditory pathway, from the outer ear and cochlea to cortical and cognitive stages. This intricate system enables humans to detect sounds across a vast dynamic range, resolve minute spectral and temporal differences, and selectively attend to a single talker in noisy, multi-talker environments. When hearing is impaired, these capacities are compromised in diverse ways, leading to a range of perceptual deficits, particularly in complex listening scenarios. While current auditory models successfully capture many aspects of normal-hearing (NH) perception, they often fail to generalise to hearing-impaired (HI) listeners. This limitation may arise because they neglect critical non-linear mechanisms, or because they lack the principled means of degrading these mechanisms in ways that correspond to the heterogeneous pathologies underlying hearing loss. Developing models that can predict individual performance in tasks related to speech intelligibility offers a powerful route to uncovering the relevant functional transformations and informing diagnostic and compensation strategies.

This thesis aims to advance efforts in computational modelling of auditory perception in both NH and HI listeners, with a particular focus on spectro-temporal modulation perception, which has been closely linked to speech perception. Within the thesis three studies are presented.

The first study introduced a revised version of the Computational Auditory Signal Processing and Perception (CASP) model. While the CASP framework has successfully predicted performance of NH listeners across a range of psychoacoustic tasks, its success in simulating individual performance of HI listeners has been limited, potentially due to simplifications in the model’s preprocessing stages. To address this, a non-linear inner hair cell (IHC) model was integrated within the model to more accurately represent the compressive input–output characteristics observed in IHC transduction, argued to be critical for speech perception. The revised model maintained predictive accuracy for NH performance across conditions of intensity discrimination, forward masking and modulation detection, while exhibiting distinct processing behaviour attributable to the non-linear IHC model. The revised model offers a more realistic foundation for simulating NH performance, as well as studying the effects of sensorineural hearing loss on the functional transformations within the model framework.

The second study applied the revised CASP model to spectro-temporal modulation (STM) detection in NH and HI listeners. STM detection thresholds have been shown to be robust predictors of speech intelligibility and have been proposed as sensitive markers of supra-threshold hearing deficits, yet, the underlying mechanisms remain incompletely understood. Behavioural data were collected from both NH and HI listeners and compared against model simulations. The revised model captured several relevant trends in the average STM threshold patterns across different temporal and spectral rates. However, it substantially underestimated group-level differences between NH and HI listeners, as well as the considerable inter-individual variability among HI listeners, suggesting that additional mechanisms are required to explain individual supra-threshold deficits.

Finally, the third study probed STM perception using a reverse-correlation paradigm applied to a STM discrimination task, in which NH listeners judged upward versus downward moving modulations. Reverse-correlation enabled the derivation of perceptual kernels that reveal how listeners weight spectro-temporal cues on a trial-by-trial basis. Comparison with a modulation filterbank model showed that group-level kernels were well-explained by the model preprocessing combined with a template-based cross-correlation strategy using only an upward-target template. However, substantial variability was observed in the data, which the model could not capture due to its fixed NH configuration. These results highlight the need to consider variability originating from different mechanisms, such as in the model’s decision-making stage. More generally, the study explored reverse-correlation as a technique for a more fine-grained evaluation of individual performance in auditory tasks, a model’s corresponding trial-by-trial predictive power and a means for constraining model parameters.

Together, the three studies demonstrate that incorporating non-linear mechanisms improves their explanatory scope, but that accounting for individual variability - arising both from peripheral and central factors - remains a major challenge. The work contributes to the development of predictive modelling frameworks that can facilitate our understanding of the transformations involved from physical stimuli to perception, with potential applications in diagnostics and the development of hearing technologies.