End-of-Turn Prediction for a Voice Agent Taking Fault Reports on a Utility Phone Line
A research poster from the School of Underlying Conditions segments 11,820 caller turns from 2,964 calls to one water utility's after-hours fault line and finds that the longest silences fall inside spoken meter numbers and street addresses rather than at the ends of turns. Replayed on held-out weeks, a slot-aware endpointer that conditions on the agent's own question and on digit-group completeness cuts off 6 per cent of meter-number readings against 31 per cent for a fixed 700 ms silence threshold, at little net cost in response latency.
| Authors | Brynjar Obiora , Vigdis Emenike , Fenna Okoro |
|---|---|
| School | School of Underlying Conditions |
| Output type | Research poster |
| Published | |
| DOI | 10.5555/slop.md27qy |
| Pages | 1 |
| Version | 1.0 |
| Licence | CC BY 4.0 |
Cites
Cite as
@misc{slop_md27qy,
author = {Brynjar Obiora and Vigdis Emenike and Fenna Okoro},
title = {Cut Off Mid-Address: End-of-Turn Prediction for a Voice Agent Taking Fault Reports on a Utility Phone Line},
year = {2026},
publisher = {Slop University},
doi = {10.5555/slop.md27qy},
url = {https://slop.university/outputs/slop-poster-end-of-turn-prediction-for-md27qy/},
version = {1.0},
note = {Research poster},
}