Enhancing Low-Resource Lampung Speech Recognition through Cross-Lingual XLSR-Wav2Vec 2.0 Pretraining

Abstract

This study investigates the application of Wav2Vec 2.0 (W2V2) and Cross-Lingual Speech Representation (XLSR) models to Lampung language speech recognition. LampungNyow v1.0 is introduced, a speech corpus designed to provide a baseline for training and evaluating Automatic Speech Recognition (ASR) for this low-resource regional language of Indonesia. The dataset enables supervised fine-tuning and standardized evaluation, addressing the lack of publicly available linguistic resources for Lampung. Several pre-trained W2V2 models on Lampung speech recognition using Word Error Rate (WER) as the evaluation metric. The evaluated models include W2V2-Base, W2V2-Large, W2V2-Large-XLSR-Indonesian, W2V2-Large-XLSR-Sundanese, W2V2-Large-XLSR-53, and the multilingual W2V2-Large-XLSR-Indonesia-Javanese-Sundanese model. Monolingual models have higher WER values, according to experimental results: W2V2-Base achieved 36,23%, while W2V2-Large achieved 36,30%. XLSR models, such as XLSR-53 (33,88%), Sundanese (33,99%), and Indonesian (33,70%), demonstrated modest improvements. The W2V2-Large-XLSR-Indonesian-Javanese-Sundanese model, which was the foundation for the Lampung automatic speech recognition system in this study, achieved lower WER of 17,39%. These findings suggest that, in contrast to more comprehensive multilingual or monolingual pretraining models, multilingual pretraining utilizing a number of Indonesian regional languages can produce acoustic and contextual speech representations that are better suited for the resource-constrained Lampung automatic speech recognition task. When compared to the baseline W2V2-Large model, the obtained WER of 17,39% indicates a relative improvement of more than 50%.