<![CDATA[The Lampung language possesses significant cultural and linguistic value, yet it remains digitally low-resource, presenting severe bottlenecks for Automatic Speech Recognition (ASR). This study addresses the technical challenges of low-resource ASR by performing a rigorous performance evaluation of fine-tuned Whisper models on a constrained local cross-dialect corpus. The dataset consists of 949 custom audio recordings (approximately 1.5 hours) capturing both Dialect A (Pesisir) and Dialect O (Abung) with manual ground-truth transcriptions. The engineering pipeline encompasses audio signal normalization, conversion to 16 kHz mono-channel, augmentation, and log-Mel spectrogram feature extraction. Three architectural scales are evaluated: Whisper-Tiny, Whisper-Base, and Whisper-Small. Evaluated via Word Error Rate (WER), Whisper-Small achieves the state-of-the-art performance with a test WER of 15.57% (validation WER 18.66%), significantly outperforming Whisper-Tiny (16.10%) and Whisper-Base (32.32%). Structural analysis reveals that Whisper-Small provides the optimal empirical balance between capacity representation and generalization stability under strict data scarcity. This research contributes a reproducible blueprint to the domain of computer science and speech processing, demonstrating that weak-supervision transformer architectures can effectively mitigate data sparsity in regional minority languages.]]>
Copyrights © 2026
