can we detect voicemail with ivr_detection with accuracy, if so how?

Thanks for reaching out, utkarshsteve7. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

so it is not supported in self hosted approach?

Thanks — you're asking if

ivr_detection

works with a self‑hosted LiveKit deployment. Short answer: It's not confirmed that

ivr_detection=True

is fully supported self‑hosted. The session option exists, but the IVR/voicemail models and inference used by

ivr_detection

appear to be deployed in LiveKit Cloud, so that behavior may be cloud‑only. What you can do instead (self‑hosted): • Implement LLM‑based voicemail detection locally to classify early call audio (LLM approaches reach ~70–80% accuracy). • Add silence detection (e.g., disconnect or flag when no user speech for ~5s after the greeting). • Combine LLM classification + silence rules for higher reliability. If you want, tell me: • whether this is outbound or inbound, • what languages/carriers you must support, • and whether you can run an LLM locally— and I’ll propose concrete implementation steps and threshold settings for a self‑hosted setup. Sources: Agent session | LiveKit Documentation | send_dtmf_events | LiveKit Documentation | How to detect and handle silence during calls | examples/voice_agents | livekit/livekit_composite | DeepWiki | examples/drive-thru | livekit/livekit_composite | DeepWiki, with interactive elements

outbound calls english openai realtime model

Thanks — this is helpful context. Quick summary:

ivr_detection

may not work in self‑hosted deployments (its models/inference appear cloud‑hosted). For outbound English calls using an OpenAI realtime model on self‑hosted infrastructure, use an LLM‑based classifier + silence heuristics. Here’s a concise plan you can implement: 1) Capture the early audio • After call connect, record or stream the first 2–6 seconds of the callee audio to the OpenAI realtime model for analysis. 2) LLM realtime classification • Send the early transcript (or short audio) to the OpenAI realtime model and ask a compact prompt to classify: human answered, voicemail/answering machine, or IVR. • Use confidence thresholds (e.g., treat >0.8 as voicemail, 0.6–0.8 as uncertain). 3) Silence & timing heuristics (combine with LLM) • If no speech detected for ~5 seconds after the greeting, flag as voicemail. • If a long prerecorded greeting pattern (monotone or TTS‑like) is detected, flag as voicemail/IVR. • Use both signals: require either LLM=voicemail OR (silence >5s AND greeting detected) to mark voicemail. 4) Retry / fallback behavior • If flagged voicemail, optionally leave message or schedule retry. • If uncertain, either attempt a short live agent callback or escalate to additional checks (longer audio sample). 5) Practical settings to try (starting point) • Audio sample: 2–6s (start at 4s). • Silence threshold: 5s. • LLM confidence cutoffs: voicemail >= 0.8, human >= 0.8; 0.6–0.8 = uncertain. 6) Observability & tuning • Log decisions, store labeled samples, and iterate thresholds per carrier/region. • Run A/B tests on sample sets to measure real accuracy and adjust. If you’d like, I can: • Provide a sample prompt for the OpenAI realtime model, • Show example code to stream audio + call the model, • Or propose exact thresholds after you share some labeled call samples. Which of those would you like next? , with interactive elements