Leivaditi, Spyretta (2023) The Role of Speech Elicitation Methods and Disease Factors in Dysartrhric ASR System Development. Master thesis, Voice Technology (VT).
|
PDF
SL_VT_thesis_with_title_corrections.pdf Download (4MB) | Preview |
Abstract
Despite significant advancements in automatic speech recognition (ASR) technology, the performance of ASR systems on dysarthric speech is still inadequate for widespread use. A reason for this is the lack of sufficiently rich and diverse dysarthric speech datasets to train machine learning models that could handle all types and varieties of such speech. Motivated by the data scarcity problem, this thesis investigates whether developers of Dutch ASR systems can take advantage of particular characteristics of dysarthric speech and increase their models' performance by selecting their training data in a strategic way. More specifically, the thesis hypothesizes a) that fine-tuning an ASR model with differently elicited speech data would lead to improved performance for the respective elicitation method, and b) that fine-tuning an ASR model with speech data affected from a specific disease would enhance model's performance on speech affected by that disease. Both hypotheses are experimentally tested by fine-tuning and evaluating a state-of-the-art self-supervised dysarthric ASR system on a new Dutch dysarthric speech dataset. The results of the experiments do not provide adequate evidence that either the elicitation method or the underlying disease of the dysarthric speakers plays a significant role in the performance of a dysarthric ASR system.
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Verkhodanova, V. |
| Date Deposited: | 26 Jul 2023 14:24 |
| Last Modified: | 26 Jul 2023 14:24 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/356 |
Actions (login required)
 |
View Item |