CARTGPT: Improving CART Captioning using Large Language Models

Communication Access Realtime Translation (CART) is a commonly used real-time captioning technology used by deaf and hard of hearing (DHH) people, due to its accuracy, reliability, and ability to provide a holistic view of the conversational environment (e.g., by displaying speaker names). However, in many real-world situations (e.g., noisy environments, long meetings), the CART captioning accuracy can considerably decline, thereby affecting the comprehension of DHH people. In this work-in-progress paper, we introduce CARTGPT, a system to assist CART captioners in improving their transcription accuracy. CARTGPT takes in errored CART captions and inaccurate automatic speech recognition (ASR) captions as input and uses a large language model to generate corrected captions in real-time. We quantified performance on a noisy speech dataset, showing that our system outperforms both CART (+5.6% accuracy) and a state-of-the-art ASR model (+17.3%). A preliminary evaluation with three DHH users further demonstrates the promise of our approach.