Krikke, Tijn (2026) The Sound of Groningen: Sub-Dialect Data Selection for Low-Resource Speech Recognition. Bachelor's Thesis, Artificial Intelligence.
|
Text
ThesisTijnKrikke.pdf Download (4MB) | Preview |
Abstract
Automatic Speech Recognition (ASR) works well for high-resource languages like English and Dutch, but low-resource languages are still a challenge. In the case of the Gronings dialect, a sub-dialect of Dutch Low Saxon, both the research and available data remain limited. To address this, a new project is collecting Dutch Low-Saxon speech data through a crowdsourced approach in which speakers record audio themselves. The current dataset includes four Gronings variants and nine other Low-Saxon variants, raising the question of whether Gronings ASR benefits more from training on Gronings data alone, on all Low-Saxon variants combined, or on a combination of the variants. To investigate this, encoder embeddings from a fine-tuned Whisper model are compared across variants using cosine similarity to identify which variants are closest to Gronings and therefore most likely to improve performance when included in training. This thesis presents the first results on this new dataset, comparing performance against base Whisper and the Gronings model introduced by Bartelds et al. (2023). Results indicate that adding more non-Gronings Low-Saxon data does not help to improve Gronings ASR, with the best WERs (0.214–0.320) across the four Gronings test sets achieved by models trained only on Gronings data.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Dhali, M.A. and Schuijlenburg, K. van |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 05 Aug 2026 06:59 |
| Last Modified: | 05 Aug 2026 06:59 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37488 |
Actions (login required)
 |
View Item |