Thesis Project Form

A central challenge in hearing research is the development of assessment procedures that better reflect the complexity of everyday communication while maintaining the control and reproducibility required in laboratory experiments.

Conventional hearing tests generally rely on simplified acoustic stimuli and highly constrained tasks. However, real-world listening often involves competing talkers, natural conversational speech, spatially distributed sound sources, visual information, and varying levels of cognitive demand. This discrepancy has motivated increasing interest in hearing-assessment paradigms with greater ecological validity.

Virtual and immersive experimental environments provide a useful framework for addressing this problem, since they allow realistic listening scenarios to be reproduced while maintaining control over the experimental conditions. An important component of such assessment paradigms is therefore the availability of speech material that is sufficiently naturalistic to resemble everyday communication but sufficiently controlled to support reproducible experiments.

The main objective of this thesis is to develop and validate a controlled naturalistic speech-on-speech corpus for hearing assessment in realistic yet experimentally controlled listening scenarios.

The work will focus primarily on the development, organization, and characterization of the speech dataset and on the implementation of a hearing-assessment test based on the developed material.

The corpus will consist of conversational speech material produced using naturalistic synthetic voices and designed specifically for speech-on-speech listening tasks.

The student will define a structured stimulus-generation pipeline, organize the resulting corpus, and characterize relevant acoustic and linguistic properties in order to assess the consistency, naturalness, and experimental suitability of the generated material.

Based on the developed corpus, a hearing-assessment test will also be implemented. During the experimental phase, participants will perform listening tasks using the developed speech-on-speech material. The experiments will assess the behavior and sensitivity of the corpus under different listening conditions. The aim will be to create a reproducible experimental procedure that can be used to assess speech perception under controlled but more ecologically relevant listening conditions.

AI-based speech synthesis will be used as a methodological tool to generate naturalistic conversational speech while allowing systematic control over relevant stimulus properties, including speaker identity, linguistic content, duration, and experimental condition.

Pupillometry will be included as a non-invasive physiological measure of cognitive listening effort. Changes in pupil diameter will be analyzed alongside behavioral performance to investigate whether variations in auditory task demand are reflected in both perceptual performance and physiological responses.

Where appropriate, spatial rendering manipulations, including individualized and non-individualized Head-Related Transfer Functions (HRTFs), may also be used during the London experiments to evaluate the sensitivity of the proposed assessment paradigm to realistic spatial listening conditions.