Thanks — you're asking how to tell whether the party who answers an outbound SIP call is an IVR (automated system) or a real human. Below are practical, implementable signals you can combine for reliable detection. Which approach do you want first — content-based (transcript/LLM) or behavior-based (silence/DTMF/turn-taking)? If you want code examples, tell me your runtime (Python/Node) and whether you have STT enabled. Quick summary of methods to use (combine for best accuracy): • LLM / transcript classification ◦ Transcribe the answered audio and run a classifier (or LLM) to detect phrases like “press 1,” “please enter,” “you’ve reached,” or robotic-sounding greetings. ◦ Trigger a predefined tool/action when a greeting matches IVR/voicemail patterns. • Silence and turn-taking detection ◦ Start a monitor when the call connects. If the callee never speaks or there’s extended silence after prompts (e.g., > 5 seconds), treat as IVR/voicemail. ◦ Track

user_started_speaking

events and measure time gaps between turns. • DTMF and menu cue detection ◦ Detect inbound/outbound DTMF tones or voice prompts that ask for keypresses (e.g., “press 1 for…”). Presence of DTMF/menu flow = IVR. ◦ Use a task like

GetDtmfTask

to actively collect and interpret DTMF if you expect IVR menus. • Audio feature and pattern detection ◦ Look for synthetic/timbral signatures (repetitive prompt audio, constant tone) or consistent prompt structure. ◦ Repetitive phrasing or exact repeated audio blocks strongly indicate IVR. Practical combined flow (recommended): 1. When the call connects, start STT streaming and a silence/turn monitor. 2. In the first 3–10 seconds: ◦ If transcript contains IVR keywords → classify as IVR. ◦ Else if DTMF detected or repeated prompt structure → classify as IVR. ◦ Else if no human reply within threshold (e.g., 5s) → run secondary checks or classify as IVR/voicemail. ◦ Else classify as human and proceed. If you want, I can: • Provide a short Python or Node sample showing STT + silence timer + simple keyword check. • Recommend thresholds (default: 5s silence threshold, 3–10s initial window) tailored to your use case. Which option do you prefer and what runtime/libraries are you using? Sources: sip | livekit/livekit_composite | DeepWiki | How to detect and handle silence during calls | Agents telephony integration | LiveKit docs | Tasks and task groups | LiveKit docs