Rollback Speech: Smart Feedback Prompts for Lost Utterances in Unstable Online Calls

We present a prototype system designed to enhance the user experience of online conferencing in the presence of poor network conditions. When connection issues cause parts of a speaker’s audio to be lost, the system uses dual automatic speech recognition (ASR) outputs—one local and one remote—to detect which utterances were not received. Upon reconnection, the speaker is provided with extracted keywords representing the information that did not reach the listener, enabling efficient and targeted repetition without full redundancy. This helps maintain conversational flow and reduces frustration during virtual communication. The system comprises real-time transcription, network degradation simulation, discrepancy detection, and lightweight keyword extraction in a modular architecture. Our “Show and Tell” demonstration allows participants to experience how the system augments communication continuity, even under conditions of packet loss or delayed transmission.