Short Description

Language is increasingly recognized as a powerful behavioural marker for Alzheimer’s disease (AD). Yet, most research so far relies on highly controlled speaking tasks and manual analyses that are impractical for clinical use. Discourse analysis, for example, has huge diagnostic potential, but the current gold-standard—manual verbatim transcription—can take 4 to 60 hours per hour of recordings.

The long-term vision is to automatically identify linguistic markers of AD and evaluating their potential to differentiate Alzheimer’s-related decline from depression-related cognitive impairment (pseudodementia). To make this clinically feasible, our goal is to develop a clinic-friendly linguistic screening tool that could be applied in real-world diagnostic routines.

The aim of this project is to build a robust automatic speech recognition (ASR) systems for spontaneous, disfluent speech. Unlike standard ASR tasks, this involves handling noisy clinical data, atypical speech patterns, and relatively small datasets, requiring innovative solutions in self-supervised learning, data augmentation and transfer learning.

Your Tasks (depending on specific project):

• Literature research
• Data preparation
• Conduction of ASR Experiments with wav2vec and Whisper
• Automatic analysis of hesitation markers
• Analysis and reporting your results (thesis writing)

Your Profile

• basic knowlegde of sound engineering and/or speech communication
• good knowledge of programming (e.g., Python)

Contact:

Barbara Schuppler (b.schuppler@tugraz.at)