Task-Based PolyLoss Improves Low-Resource Speech Recognition

Multilingual meta learning has emerged as a promising paradigm for transferring the knowledge from source languages to facilitate the learning of low-resource target languages. Loss functions are a type of meta-knowledge that is crucial to effective training of neural networks, however, the misalignment between the loss functions and learning paradigms of meta learning degrade the network performance. To address this challenge, we propose a new method called Task-based Meta PolyLoss (TMPL) for meta learning. By regarding speech recognition tasks as normal samples and apply PolyLoss to meta loss function, TMPL can be denoted as a linear combination of polynomial functions based on task query loss. We conduct extensive experiments on four low-resource languages from the IARPA BABEL dataset. The results show a significant improvement in performance when applying our TMPL to meta learning.