Stuck on a persistent issue where the Voice AI transcription confidence scores are significantly lower for calls routed through our Singapore BYOC trunks compared to the standard Genesys Cloud trunks. We are observing a drop in intent detection accuracy specifically when the SIP RTT exceeds 200ms, which is common during peak hours on our secondary carrier failover path.
The environment involves 15 BYOC trunks configured with strict outbound routing rules. When the primary trunk fails, traffic shifts to the secondary carrier, introducing an average latency of 250ms. In these scenarios, the Voice AI agent frequently fails to recognize standard phrases, resulting in a critical drop in customer satisfaction metrics. The Architect flow is configured to trigger a fallback to human agents after three failed intent matches, but this happens prematurely due to the low confidence scores rather than actual user confusion.
I have verified the SIP registration status and confirmed that the audio quality (MOS score) remains above 4.0, indicating that packet loss is not the primary culprit. However, the transcription logs show fragmented words and missing syllables, suggesting that the audio buffering logic within the Voice AI module might be sensitive to the specific jitter patterns introduced by the carrier gateway.
Has anyone encountered similar issues with Voice AI performance on high-latency BYOC connections? I am looking for specific configuration adjustments in the Architect flow or Voice AI settings that can mitigate this sensitivity. We cannot simply increase the timeout values, as that would degrade the overall customer experience. Any insights into how the Voice AI engine handles jittery audio streams from external trunks would be greatly appreciated. Specifically, I need to know if there is a way to tune the confidence threshold dynamically based on the trunk’s real-time latency metrics.