Ontology Memory-Augmented ASR Correction for Long Text-Speech Interleaved Conversations

The paper introduces an ontology memory-augmented framework for automatic speech recognition (ASR) correction tailored for long text-speech interleaved conversations. This framework utilizes a dynamically updatable ontology memory to organize interaction history, enhancing the retrieval of contextual evidence for ASR corrections. Evaluated on the RAMC-Corr dataset, the method demonstrates significant improvements in correction accuracy across various backbone settings, highlighting its potential for more effective context-aware ASR systems.