Iterative pseudo-labeling methods for improving speech recognition

Abstract

In recent years, pseudo-labeling methods can reduce the difficulty of building speech recognition systems, in end-to-end automatic speech recognition (ASR). Iterative pseudo-labeling (IPL) is a classical semi-supervised algorithm that can efficiently perform multiple pseudo-labeling iterations on unlabeled data as acoustic models evolve. We incorporate the language model to generate pseudo-labeling based on IPL using the language model for decoding and data augmentation, and make new attempts on the selection of pseudo-labeling. The effectiveness of the improved approach is demonstrated by simulating low resources and standard settings and obtaining a word error rate better than IPL on the LIBRISPEECH test.