Dictionary-driven Chinese ASR Entity Correction with Controllable Decoding

The mainstream ASR error correction system is mainly based on the encoder-decoder structure, by learning the mapping of incorrect text to correct text. Although this approach showed good results in ASR error correction, it mainly solves the spelling errors that occur at high frequencies, while struggling with low-frequency ASR entity errors, especially those entities unseen in the training set. Another downside of the structure is that it may produce undesirable hallucination entities. In this paper, we introduce a DiCODER model for Chinese ASR entity error correction. Specifically, we first retrieve relevant entities from the entity dictionary based on the similarity of the text and phonetic information and then fuse them with ASR text for joint encoding. Next, a controlled decoding strategy utilizing dynamic vocabulary is proposed to generate more reasonable results. Experimental results on three publicly available datasets demonstrate the effectiveness of our proposed method.