Improving low-resource ASR using bilingual fine-tuning with language identification: a cross-linguistic evaluation
Researchers have developed a method for improving automatic speech recognition (ASR) in low-resource languages through bilingual fine-tuning. The study evaluated this technique across nine diverse language pairs, using language identification tokens to distinguish between languages during training and inference. Results indicate that bilingual fine-tuning is effective when language identification is accurate, and providing the identification token at inference further boosts performance in cases of lower accuracy. AI
IMPACT This research offers a method to improve speech recognition for languages with limited data, potentially increasing accessibility and usability of AI technologies globally.