BROWSE
Detail
Conference Paper
Considerate Listener Modeling for Korean Streaming Backchannel Prediction
Cited - time in 
Share 
Share
- Authors
- Yong-Seok Choi, Seung Hi Kim, Sung Yup Lee
- Issue Date
- 2026-09
- Citation
- International Speech Communication Association (INTERSPEECH) 2026, pp.3405-3409
- Publisher
- ISCA
- Language
- English
- Type
- Conference Paper
- Abstract
- Mistimed backchannels can interrupt ongoing speech and degrade user experience. A Considerate Listener concentrates backchannel interventions near pause-proximate regions in real time, since backchannels cluster around speaker pauses. However, not all pauses invite a backchannel. Pauses that follow utterances calling for a reply may be pragmatically inappropriate for backchannel intervention. Acoustic cues alone may not capture the linguistic context needed to distinguish these cases, so acoustic-only predictors tend to over-predict. We therefore propose a zero look-ahead streaming backchannel predictor combining pause-aware soft scaling with Q-Former fusion of partial ASR hypotheses. We introduce Semantic False Discovery Rate (Semantic FDR) to quantify backchannel predictions in pragmatically inappropriate pause regions. On a Korean counseling corpus, our approach reduces Semantic FDR by 53.8% and improves Macro-F1 by 2.16 pp compared to an acoustic-only baseline.
- Keyword
- streaming backchannel prediction, considerate listener, pause-aware soft scaling, text-guided fusion, semantic false positives
- KSP Keywords
- Acoustic cues, Look-ahead, User Experience, false discovery rate, false positive, real time