Lankheet, Amber (2025) A Cross-Lingual Approach to Dutch Dysarthric Speech Recognition. Master thesis, Voice Technology (VT).
|
PDF
MSc4367073ALankheet.pdf Download (587kB) | Preview |
Abstract
This thesis explores how well a multilingual self-supervised speech recognition model, XLSR-53, can understand Dutch speech from people with dysarthria, a motor speech disorder that affects pronunciation. Automatic Speech Recognition (ASR) can help people with dysarthria communicate more easily, but current systems often fail because of unclear or unusual speech patterns. A common idea in recent research is that using data from many languages (cross-lingual training) might help models better handle this kind of variation. To test this, I compared four setups: using a high-resource Dutch model without extra training, fine-tuning on healthy Dutch speech, fine-tuning on English dysarthric speech, and a combination of both. I evaluated each model’s performance using Word Error Rate (WER) on Dutch dysarthric test data. Although none of the fine-tuned models outperformed the high-resource baseline, the combined approach did slightly better than the models fine-tuned on only one type of data. The findings show that fine-tuning with mismatched or limited data can make performance worse, even when using advanced models. This research gives insight into what does and doesn’t work for dysarthric speech recognition. It also highlights important issues, such as limited speaker diversity and age differences in the data, and suggests future research could focus on phoneme-level evaluation and training specific parts of the model to improve results. Overall, this work helps researchers and developers better understand how to create more inclusive ASR systems that support people with speech impairments.
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Verkhodanova, V. |
| Date Deposited: | 16 Jun 2025 11:06 |
| Last Modified: | 16 Jun 2025 11:06 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/659 |
Actions (login required)
 |
View Item |