Developing an English-to-Indonesian speech-to-text as a foundation for Sasak language translation using the mBART algorithm

The advancement of speech-to-text technology has revolutionized data processing, enabling direct voice input, accelerating data entry, and supporting applications like auto-subtitling. However, challenges in translation accuracy persist, particularly for large datasets and regional languages. This study introduces an mBART-based translation model to convert English speech into Sasak text, effectively addressing dialectal variations. The model incorporates non-standard Indonesian as an intermediary, producing more natural and contextually accurate Sasak translations. Evaluated using the BLEU metric, the fine-tuned model achieved a score of 79.78 compared to 47.99 for the standard model, while the implementation of the CA-FIT Translation framework further improved performance, increasing BLEU scores from 48.79 to 64.99. These results highlight the effectiveness of fine-tuning and advanced frameworks in improving translation quality and preserving regional languages, supporting the development of more accurate multilingual speech-to-text technology.