audio-commentary

This paper presents a low-latency real-time audio game commentary system that uses LLM-based parallel text generation to reduce inter-utterance silence from 9.6 to 0.3 seconds, significantly improving perceived speaking rhythm compared to sequential baselines.