Reproduce “Towards Pretraining Robust ASR Foundation Model with Acoustic-Aware Data Augmentation” on OpenResearch

About this paper

Dancheng Liu, Amir Nassereldine, Chenhui Xu, Jinjun Xiong

Researchers from the University at Buffalo empirically demonstrate that acoustic diversity is a more critical factor for ASR model robustness and generalization than linguistic diversity alone, achieving up to 20.63% relative Word Error Rate (WER) reduction on accented speech and 18.4% on children's speech through novel acoustic-centric data augmentation.

Key insights

• Linguistic diversity in training data, without corresponding acoustic variation, is insufficient for an ASR model to generalize to real human speech, leading to complete recognition failure.
• Acoustic diversity is empirically a more critical factor than linguistic diversity for achieving robustness and generalization to out-of-distribution (OOD) speech in ASR foundation models.
• Conventional data augmentation methods like SpecAugment primarily influence linguistic inference rather than directly enriching the acoustic diversity needed for OOD generalization.

Summary by alphaXiv.