Institutional Repository of Vilnius University
Šnekos atpažinimo modelio papildomo apmokymo strategijų lietuvių kalbai tyrimas /
Abstract
dc:descriptionThis study analyzes the adaptation of the multilingual Whisper large‑v3 model for Lithuanian spontaneous speech recognition. The LIEPA‑2 spontaneous speech data are used, covering speech variation, different recording conditions, and dialectal features found in naturally spoken Lithuanian. Lithuanian is examined as a low‑resource language, for which large‑scale, publicly available, and high‑quality annotated speech datasets remain limited. The aim of the study is to investigate different model fine‑tuning strategies: full fine‑tuning, selective fine‑tuning by freezing part of the model layers, and low‑rank adaptation‑based finetuning. The study also analyzes how the performance of each strategy depends on the selected hyperparameters. The final models are compared using the same one‑epoch training budget. For this reason, the conclusions are interpreted within the limits of this experimental setup. In the final analysis, model quality is evaluated using word error rate, character error rate, semantic similarity, phoneme edit distance, keyword recall, and exact‑match rate. The result of the base model without fine‑tuning is also included to assess how much each adaptation strategy improves transcription quality.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Ezerskis, Juras,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308336823&prefLang=en_US