Ontology Memory-Augmented ASR Correction for Long Text-Speech Interleaved Conversations
Researchers have developed a novel ontology memory-augmented framework to improve automatic speech recognition (ASR) correction for long, interleaved text and speech conversations. This approach organizes interaction history into a dynamic ontology memory, storing entities, terminology, and semantic relations as retrievable nodes. Experiments using the newly constructed RAMC-Corr dataset demonstrate that this method significantly enhances correction accuracy compared to direct correction methods. AI
IMPACT This new ASR correction framework could improve the accuracy of speech-to-text systems in complex, long-form conversational contexts.