Reproduce “Optimizing Estonian TV Subtitles with Semi-supervised Learning and LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Artem Fedorchenko, Tanel Alumäe
Researchers from Tallinn University of Technology developed an automated system for generating high-quality, same-language Estonian TV subtitles by fine-tuning the Whisper ASR model, applying iterative pseudo-labeling, and integrating Large Language Model post-editing. The optimized approach achieved a Subtitle Edit Rate (SubER) of 33.1, demonstrating improved subtitle quality and style over verbatim transcripts.
Key insights
- Models explicitly fine-tuned on professional subtitle data (SubER 35.1) perform substantially better for subtitle generation than those trained on verbatim transcripts (SubER 51.5), emphasizing the importance of domain-specific training.
- Iterative pseudo-labeling consistently improves subtitle quality by effectively leveraging unlabeled data for less-resourced languages (SubER decreased from 35.1 to 33.4 after two IPL iterations).
- Large Language Model post-editing is most effective when applied at test time to the final subtitle outputs (improving SubER from 33.4 to 33.1) rather than for refining pseudo-labels during the training phase.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.