Multilingual Meta-Transfer Learning for Low-Resource Speech Recognition

Abstract

This paper proposes a novel meta-transfer learning method to improve automatic speech recognition (ASR) performance in low-resource languages. Nowadays, we are witnessing high interest in low-resource ASR tasks aiming at delivering feasible and reliable systems with very limited data. The main challenge is the design and development of a methodology to address the issue of data scarcity. Our proposed meta-transfer learning approach combines two well-known machine-learning methods: transfer learning and meta-learning. We propose their integration that can ameliorate the training bottlenecks and overfitting issues with pre-training models on low-resource speech data. For evaluation, we conduct extensive multilingual ASR experiments on the Common Voice Corpus and Globalphone Corpus and compare the performance of the meta-transfer learning, meta-learning, and transfer learning methods. The proposed meta-transfer learning achieves a relative character error rate (CER) reduction of 11.62% over meta-learning and a relative CER reduction of 10.86% over transfer learning in low-resource experiments. We used less than 15 minutes of data for each target language in near-zero resource language experiments. Our meta-transfer learning approach achieved an average CER of 25.25% less than meta-learning and transfer learning. These results clearly demonstrate that the proposed integration works well in ASR tasks in languages with very limited data resources.