EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Optimization of Data Pre-Processing to Improve the Accuracy of Sundanese Speech Recognition Using LLM GPT-4

Abdul Halim Anshor, Amali, Sugeng Budi Rahardjo, Ahmad Fauzi, Tri Ngudi Wiyatno

International Journal on Advanced Science, Engineering and Information Technology · 15.10.2026 · 16 (5) · 1636-1642

DOI: 10.18517/ijaseit.16.5.21083 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal on Advanced Science, Engineering and Information Technology, online erschienen am 15.10.2026. Verfasst von Abdul Halim Anshor und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Although multilingual speech recognition systems have made notable progress, transcription quality remained inconsistent when processing Sundanese speech. During a preliminary review of the generated transcripts, errors repeatedly appeared in personal names, infrequent lexical items, and expressions whose interpretation depended heavily on context. These errors often altered an utterance's intended meaning and limited the transcripts' practical usefulness. These errors persisted even with a modern recognition model, making the resulting transcripts less reliable for practical use. To examine whether we could reduce these errors without modifying the speech recognizer, we used GPT-4 as a post-editing component for transcripts generated by Whisper. The evaluation involved 827 Sundanese utterances obtained from a publicly available corpus. Whisper first transcribed each utterance, then GPT-4 revised it using contextual examples in a few-shot prompting scheme. We measured performance using Word Error Rate (WER). The original transcripts contained many recognition mistakes, yielding an average WER of 1.024. After post-editing, the average WER fell to 0.802. Improvements were observed most often in fragmented words, incorrectly recognized names, and lexical forms that depended strongly on context. These results indicated that contextual reasoning available to a large language model complemented information extracted from the acoustic signal. The study showed that post-editing with GPT-4 improved transcription quality and provided a practical alternative for enhancing Sundanese speech recognition without additional speech annotation or acoustic model retraining.

Zitieren

APA

Anshor, A. H., Amali, Rahardjo, S. B., Fauzi, A., & Wiyatno, T. N. (2026). Optimization of Data Pre-Processing to Improve the Accuracy of Sundanese Speech Recognition Using LLM GPT-4. International Journal on Advanced Science, Engineering and Information Technology, 16(5), 1636-1642. https://doi.org/10.18517/ijaseit.16.5.21083

Vancouver

Anshor AH, Amali, Rahardjo SB, Fauzi A, Wiyatno TN. Optimization of Data Pre-Processing to Improve the Accuracy of Sundanese Speech Recognition Using LLM GPT-4. International Journal on Advanced Science, Engineering and Information Technology. 2026;16(5):1636-1642. doi:10.18517/ijaseit.16.5.21083

BibTeX
@article{Anshor2026,
  author = {Anshor, Abdul Halim and Amali and Rahardjo, Sugeng Budi and Fauzi, Ahmad and Wiyatno, Tri Ngudi},
  title = {Optimization of Data Pre-Processing to Improve the Accuracy of Sundanese Speech Recognition Using LLM GPT-4},
  journal = {International Journal on Advanced Science, Engineering and Information Technology},
  year = {2026},
  volume = {16},
  number = {5},
  pages = {1636--1642},
  doi = {10.18517/ijaseit.16.5.21083}
}
Kostenlos starten →