<![CDATA[Automatic speech recognition (ASR) for regional languages is often constrained by limited paired speech–text data and by evaluation settings that can hide important differences between familiar and previously unseen transcript content. This study examines full-model fine-tuning of Whisper Small for Makassar ASR using a processed corpus of 6,119 utterances (6.3882 h). The corpus was partitioned into 4,894 training utterances, 612 validation utterances, and a locked 613-utterance test set. Because the same three speakers occur in all partitions, the evaluation is explicitly speaker-dependent. For a fair reviewer-driven baseline, the original zero-shot Whisper Small model and the fine-tuned checkpoint were re-evaluated on the identical test items with the same processor, 16-kHz audio preparation, Indonesian transcription setting, deterministic one-beam decoding, and a shared 225-token generation safety cap. Under this controlled protocol, overall WER decreased from 120.84% to 51.78%, while CER decreased from 62.00% to 16.19%. A post-hoc transcript-content audit identified 136 seen-content test utterances (22.19%) and 477 unseen-content utterances (77.81%). On unseen content, WER decreased from 125.88% to 60.91%, an absolute reduction of 64.97 percentage points. Error decomposition showed that substitutions remained the dominant error class after fine-tuning. The zero-shot system also showed repetitive long-generation behavior in targeted ceiling-hit cases. These findings support language- and corpus-specific adaptation of Whisper Small for Makassar, while also showing why aggregate error rates should be interpreted together with content-overlap, speaker-design, decoding, and reproducibility audits.]]>
Copyrights © 2026
