JSPG: Dynamic Dictionary Filtering via Joint Semantic-Pinyin-Glyph Retrieval for Chinese Contextual ASR

In three linesJSPG introduces a dynamic filtering framework for Chinese contextual ASR combining semantic, pinyin, and glyph features. The approach uses an extended Smith-Waterman algorithm to score N-best hypothesis sequences against keywords. Experiments on Aishell-1 and RWCS-NER datasets show significant improvements in keyword recognition accuracy.

Summary generated by Claude — human-verified