This thesis presents a novel evaluation framework designed to assess the performance of automatic speech recognition (ASR) systems when applied to speech affected by aphasia. The study begins with a comprehensive review of existing literature, identifying key challenges in applying ASR technologies to disordered speech and highlighting gaps in current evaluation methods. Building on traditional error rate metrics such as Word Error Rate (WER) and Character Error Rate (CER), the proposed framework introduces a more granular analysis tailored to the specific phenomena of aphasic speech, including paraphasias, semantic distortions, and disfluencies. Using large annotated datasets and state-of-the-art ASR models, the framework was applied in empirical testing to reveal model-specific weaknesses. Results show that while general-purpose ASR systems achieve acceptable WER on non-aphasic speech, their performance degrades significantly on aphasic speech samples, particularly in transcribing disfluencies such as filled pauses. Overall, this research offers valuable tools for both researchers and clinicians aiming to improve ASR applications in aphasia treatment and assessment.
| Date | 9 Sept 2025 |
|---|
| Original language | French |
|---|
| Awarding Institution | - École de technologie supérieure
|
|---|
| Supervisor | Sylvie Ratté (Supervisor) & Pierre André Ménard (Co-supervisor) |
|---|
Dupuis Desroches, J. (Author),
Ratté (Supervisor) & Ménard (Co-supervisor),
9 Sept 2025Student thesis: Master's thesis › Master in Engineering: Information Technology Engineering