Wang, Yi (2026) Language Model Integration and Confidence-Based Self-Training for Low-Resource Gronings Speech Recognition. Bachelor's Thesis, Artificial Intelligence.
|
Text
final-2.pdf Download (688kB) | Preview |
Abstract
Automatic speech recognition (ASR) performs poorly for low-resource language varieties such as Gronings, a Low-Saxon variety spoken in the north of the Netherlands. Building on Bartelds et al. (2023), this study investigates two routes to improving a fine-tuned XLS-R acoustic model without new transcriptions: (i) integrating an external n-gram language model at decoding time, and (ii) confidence-based self-training, filtering low-confidence pseudo-labels while scaling the unlabelled pool with new out-of-domain audio. A KenLM model trained on 2.5 million words of written Gronings yields a consistent relative WER reduction of roughly 10 to 12 percent on both an in-domain and an out-of-domain test set, persisting even for a much stronger acoustic model. For self-training, confidence filtering of the limited in-domain pool does not help and can hurt performance, since filtering trades quantity for only marginal quality gains. Adding a moderate amount (3 to 10 hours) of new unlabelled audio lowers WER on both test sets, but adding the full 22 hours of noisy audio shows diminishing returns and even reversal. On this large, noisy pool, the outcome reverses: every filtered model beats the baseline out-of-domain, the best result obtained by keeping the top 75 percent, while in-domain performance stays close to baseline. This shows that on a large, noisy pool, confidence filtering can meaningfully improve out-of-domain generalisation without hurting in-domain performance.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Schuijlenburg, K. van and Dhali, M.A. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 05 Aug 2026 06:57 |
| Last Modified: | 05 Aug 2026 06:57 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37487 |
Actions (login required)
 |
View Item |