We need help with outbound direct-live pickup behavior in Retell.
Our IVR mortgage agent is separate and should not be changed. The issue is only with a separate direct-live agent for lenders that answer with a live representative immediately.
Agent:
agent_d63900fb287b3d8e910bb3dc2a
Problem:
On direct-live outbound calls, the first live representative pickup is failing. Reps say “hello?” repeatedly or the first agent audio is clipped/delayed, then the rep hangs up within 10-25 seconds. Prompt changes have not fixed it.
Failed call IDs:
call_a9b35b10792076fd8af0ca6ea4a
call_60c45fc6861b68b6e2b58d7f570
call_3ab7afaa838939e6fd97f0b2a88
call_4f7f43347af728c450bd78473fb
Observed behavior:
- First Trust: rep answered, said hello multiple times, agent responded late, rep hung up.
- CorTrust: agent first word clipped, rep said hello, agent completed opener, rep hung up.
- Earlier versions had similar delayed/clipped first-response behavior.
Previous direct-live settings tested:
start_speaker: user
begin_after_user_silence_ms: 1800
responsiveness: 0.55
interruption_sensitivity: 0.25
New test version published:
start_speaker: agent
begin_message_delay_ms: 900
begin_after_user_silence_ms: null
responsiveness: 0.72
interruption_sensitivity: 0.25
Ask AI suggested using start_speaker: agent with begin_message_delay_ms, and possibly increasing interruption_sensitivity from 0.25. We need Retell support to review the actual calls and confirm recommended settings for outbound calls where a live receptionist answers immediately.
Questions:
- Should direct-live pickup use start_speaker: agent with begin_message_delay_ms, or user-first with begin_after_user_silence_ms?
- What responsiveness/interruption_sensitivity settings do you recommend for this exact live receptionist pickup case?
- Could this be voice/telephony setup rather than prompt/settings?
Hello @tnussbeck Checking with the team on this. Regards
Hello @tnussbeck we pulled all of them and the transcripts/latency telemetry tell a clear story.
What’s actually happening on turn 0 (same pattern on all four calls, v7 and v8):
- Rep finishes their greeting around 3.1–3.9 s into the call.
- Agent’s first word (“Hi,”) lands at 6.0–11.1 s — i.e. 2.8–4.4 s of dead air after the rep stops speaking.
- LLM first-token (~555 ms), TTS first-chunk (~189 ms, Sloane voice), ASR (~216 ms), turn-taking (~16 ms) are all healthy. The remaining ~1.8 s residual matches
begin_after_user_silence_ms=1800 exactly.
- The opener is also delivered as two sentences with a 1.2–1.5 s pause between them, which adds to the “slow / clipped” perception.
- No first-word truncation in any transcript; no error codes; no TTS fallback events. This is not a voice or telephony issue.
Why the v8 change didn’t help: start_speaker and the begin-message live on the conversation flow’s begin node, not on the agent root. On both v7 and v8 the flow still has startSpeaker: user and no begin_message on the start node, so runtime behavior is identical to v7 regardless of what you set on the agent record.
Recommended config for direct-live pickup — pick one of two patterns:
- Agent-first (cleanest for live pickup): edit the conversation flow’s begin node — set
startSpeaker: agent, add an explicit one-sentence begin_message (e.g. “Hi, this is Emily with Aspen Contracting — calling about a loss-draft check, do you have a moment?”), and set begin_message_delay_ms to 500–800 ms so it lands shortly after the rep’s “hello?”. Republish the flow.
- User-first (current pattern, just faster): keep
startSpeaker: user but drop begin_after_user_silence_ms from 1800 → ~800–1000 ms. That removes ~1 s of dead air after the rep stops talking.
In both patterns:
- Collapse the opener to a single sentence to eliminate the 1.2–1.5 s mid-utterance pause.
- Keep
responsiveness: 0.55 — it’s well-placed for live-rep pickup.
- Raise
interruption_sensitivity from 0.25 → 0.4–0.6 so brief rep interjections aren’t talked over.
- Keep
enable_backchannel: false, and don’t add ambient_sound.
Thank You