Final Year Project

Project Title

YAARAN - Speech-to-Text Translator for Burushaski

Description

YAARAN develops the first speech-to-text translation system for Burushaski, an endangered language isolate spoken in northern Pakistan. We created a parallel corpus of 7,115 Burushaski audio samples paired with English translations from 42 speakers, representing ~10 hours of speech data. We fine-tuned three Whisper model variants (Small: 244M, Medium: 769M, Large: 1.55B parameters) to systematically evaluate scaling benefits for low-resource translation. Using normalized evaluation metrics, we achieved BLEU 70.90 and 75% word accuracy, with Whisper-Medium offering an optimal performance-efficiency balance (BLEU 69.94). Our work establishes the first speech translation benchmark for Burushaski, demonstrates clear but diminishing returns with model scaling, and provides a replicable methodology applicable to other endangered languages worldwide.