The Genesys AI NLU engine is returning intent confidence scores below 0.3 for voice interactions routed through our Singapore BYOC trunks, despite identical text inputs scoring above 0.85. This discrepancy appears only when the audio path traverses the Bandwidth SIP leg. The Architect flow uses the standard ‘Get AI Assistant Intent’ data action.
Is there a known latency or codec issue affecting speech-to-text accuracy on specific carrier failovers?
This happens because the speech-to-text engine receiving a different audio format than expected due to the Bandwidth SIP leg negotiation. The default codec settings for BYOC trunks often default to G.729, which degrades transcription accuracy significantly compared to G.711 (ULAW/ALAW).
Check your BYOC trunk configuration in the Admin portal. Ensure the “Preferred Codecs” list has G.711 at the top. If G.711 is disabled or lower priority, the STT engine struggles with the compressed audio, dropping confidence scores. I see this often when agents are on mobile networks that force codec renegotiation. A quick test is to force G.711 on the trunk and rerun a few calls. If confidence jumps back above 0.8, that is your fix. Also, verify the “Transcription Language” matches the exact region variant (e.g., en-SG) in the AI Assistant configuration. Mismatched locale settings can also cause drift.