Extending Pronunciation Dictionary with Automatically Detected Word Mispronunciations to Improve PAII's System for Interspeech 2021 Non-Native Child English Close Track ASR Challenge

Abstract

This paper proposed to automatically detect mispro-nounced words over the regions that have low Goodness-of-Pronunciation scores through a constrained phone decoder, then add these word mispronunciations into the orthodox lexicon without colliding with existing pronunciations, finally use the expanded lexicon for decoding non-native speech. The constrained phone decoder is compiled by using a phone-level automatically generated one-edit-distance network to eliminate the need of extended recognition networks designed by phonologists. Results and analysis have shown that the pro-nunciation dictionary extension is effective in improving WER performance for non-native speech recognition. This paper also described the details of PAII’s single-pass fusion-free hybrid system for this Interspeech 2021 non-native children English close track ASR challenge, especially showed the effective use of non-speech segments in the training set as noise sources to perform noise augmentation on the training data, and also conducted a comparision of acoustic models with different neural network architectures with analysis. Final WERs of 12.10%/28.25% are obtained compared to a well-optimized baseline with WERs of 13.37%/33.51% on development/evaluation set, respectively.