Meta-learning for Indian languages: Performance analysis and improvements with linguistic similarity measures

Indian languages share a lot of overlap in acoustic and linguistic content. Though different use writing systems, the phoneme sets logically overlap. Most these are low-resourced, lacking enough annotated speech data to build good automatic recognition (ASR) systems. Recently proposed model-agnostic meta-learning (MAML) algorithm has shown great success fast adaptation multilingual models unseen datasets. In this work, we establish usefulness MAML pretraining quickly building reasonably ASRs for low-resource languages. significantly outperforms joint training its capability few-shot learning faster adaptation. On average, yields absolute improvements 5.4% CER 20.3% WER over fast-adaptation setting with five epoch fine-tuning. Further, exploit similarities source transcriptions target through loss-weighing scheme during improve performance models. Similarity-based loss-weighings yield 0.2% 1% on average.