Large Language Model Fine-Tuning for Indonesian-to-Sundanese Speech-to-Speech Translation Based on Undak-Usuk Basa

Abstract

The Sundanese language is a vital cultural pillar in West Java, yet the practice of Undak-Usuk Basa (sociolinguistic speech levels) has experienced significant degradation among the younger generation. Conventional Neural Machine Translation (NMT) models often struggle to capture sociolinguistic context, frequently resulting in a heavy bias toward informal speech levels (Basa Loma). This study develops a speech-to-speech translation (S2ST) prototype from Indonesian to Sundanese utilizing a cascaded architecture consisting of Automatic Speech Recognition (ASR), a Large Language Model (LLM), and Text-to-Speech (TTS). We leverage Meta's Llama 3 (8B Instruct) as the core language translator, fine-tuned with Quantized Low-Rank Adaptation (QLoRA). To address the extreme low-resource constraints of the Sundanese language, our training begins with a high-fidelity seed dataset containing 881 manually curated parallel sentence pairs across Loma, Sedeng, and Lemes levels (representing 2,643 initial instruction pairs). Data collected from digital sources was asynchronously validated by certified Sundanese educators. This seed corpus was subsequently expanded to a total of 6,843 instructions through synonym substitution and back-translation techniques. Standard evaluation metrics demonstrate that the proposed model achieves an overall BLEU score of 66.42 and a METEOR score of 86.77%. Qualitative analysis indicates that METEOR is a more representative metric for low-resource honorific languages like Sundanese because it accounts for lexical synonyms and morphological stemming where exact n-gram matching underperforms. Human evaluation across three equal demographic cohorts (children, juveniles, adults/linguists) yields a mean score of 3.42 out of 4.00 (“Excellent / Highly Natural”).