Slide-EC: A Slide-Based Error Correction Framework for Speech Recognition

Conventional automatic speech recognition (ASR) systems frequently face performance constraints, particularly in environments with an abundance of domain terminology. We introduce an innovative ASR error correction system that integrates slide information. In this paper, we propose Slide-EC, an enhanced error correction method that combines acoustic data with external visual representations to produce precise speech transcription contexts. Our approach employs optical character recognition (OCR) and named entity recognition (NER) technologies to capture entity information from slides. Additionally, we leverage the international phonetic alphabet (IPA) representations of these extracted entities to search and rectify errors in ASR outputs. Experiments conducted on two different language technical lecture datasets, each equipped with ground truth transcriptions, demonstrate that our method consistently enhances ASR performance. All in all, the Slide-EC system not only autonomously constructs an entity database from slide information but also can support asynchronous correction for streaming ASR.