asr-errors

This paper presents PiDA, a phonetically-informed data augmentation method for Vietnamese speech translation that improves robustness by generating ASR-like corruptions using phonetic word embeddings, achieving up to +2.04 BLEU on noisy outputs.