For Slovene, a less-resourced language, automatic speech recognition remains a challenging task, while modern large language models, with their rich linguistic knowledge, open new possibilities for improving its accuracy. In this thesis we systematically study and evaluate four groups of methods for integrating large language models into a Slovene speech recognition system, ranging from shallow integration (hypothesis rescoring, prompting, and hypothesis-to-transcript mapping) to deep integration (the multimodal SALM model). We tested the methods with eight large language models and compared them on six Slovene evaluation sets using word error rate (WER) and inference speed (RTFX). The best accuracy is achieved by rescoring with the SlovenianGPT model (average WER 10.26 %). We found that for Slovene speech, large language models are most effective when they use speech recognition hypotheses and we strictly define how they are used or fine-tune them for the task. They therefore perform best in hypothesis rescoring or fine-tuned hypothesis-to-transcript mapping, but not in unconstrained transcript generation. The analysis of computational cost shows that all methods involving large language models are considerably slower and more memory-intensive than the baseline recognizer, with hypothesis-to-transcript mapping offering the best trade-off between accuracy and speed among them.
|
