Abstract
Conversational voice systems that use fixed silence timeouts for voice activity detection (VAD) may prematurely interrupt users who have speech dysfluencies or natural hesitations, potentially leading to incomplete inputs and erroneous system responses. Techniques are described for analyzing a mirrored audio stream in a parallel, out-of-band process to detect acoustic patterns indicative of speech dysfluency, for example, without adding substantial latency to the live conversation. Based on this real-time analysis, a calibration engine can calculate and apply an extended silence timeout for the user's active session. This dynamic, in-session adjustment of the VAD parameter may improve the likelihood that a complete user utterance is captured, which can result in more accurate input processing and a more appropriate system response.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Raman, Sundarrajan, "Dynamic Calibration of Voice Activity Detection for Conversational Systems", Technical Disclosure Commons, (October 12, 2026)
https://www.tdcommons.org/dpubs_series/12096