Three presets and the values behind them: when a turn ends, when a caller can interrupt, and what happens on a silent line.
Three settings decide the rhythm of a call: how sensitive the agent is to speech, how long a pause has to be before it takes its turn, and how much speech a caller needs to produce to cut the agent off. Three more decide what happens when nobody says anything at all.
All six live on the Call handling tab, in the Turn-taking & silence card. Its first line says what the current values do on a call, for example "Answers about 0.7 s after the caller stops talking; the caller can cut in after 0.05 s of speech. If the caller never speaks, it hangs up after 1 min. Once they have spoken, it checks in after 2½ min of silence and hangs up at 3 min." It changes as you edit. Most agents should use a preset.
agents:voice:manage permission.Call handling → Turn-taking & silence → Quick presets. Clicking one writes all six values at once; the active preset is highlighted with a tick while the values still match it. The three turn-taking numbers sit under Adjust the numbers, closed while a preset matches and open when your values are your own.
| Preset | For | Activation Threshold | Min Silence | Min Speech | Caller never speaks | Check in after | Hang up after |
|---|---|---|---|---|---|---|---|
| Conversational — Balanced for everyday calls | Most agents. The speech detector's own defaults. | 0.5 | 0.55 s | 0.05 s | 60 s | 150 s | 180 s |
| Fast-paced — Snappy turn-taking | Short transactional calls, confirmations, quick lookups | 0.3 | 0.35 s | 0.03 s | 30 s | 90 s | 120 s |
| Patient (IVR) — Tolerates long pauses | Callers reading a number off a document, navigating a menu, or thinking | 0.6 | 0.75 s | 0.1 s | 90 s | 240 s | 300 s |
A new agent starts on Conversational.
All three are in seconds. The form's help text under each field says the same thing.
Activation Threshold — voice detection sensitivity, between 0 and 1. How confident the agent must be that it is hearing a human voice rather than background noise. Lower is more sensitive: it picks up quiet speech, and also picks up traffic, a television and a colleague two desks away. Raise it for callers on speakerphone in loud places; lower it for soft-spoken callers on a clean line.
Min Silence Duration — the pause that ends the caller's turn. The agent waits this long after speech stops before treating the turn as finished and starting to answer. This is the single biggest lever on how the call feels. Too short and the agent talks over someone who was mid-thought; too long and it seems slow.
Min Speech Duration — speech needed before the caller can interrupt. How much continuous speech the caller must produce for the agent to stop talking and listen. Low values let a "yes", "no" or "wait" cut the agent off; higher values mean a cough, a "mm-hm" or a door closing will not.
flowchart LR
A["Caller speaks"] --> B{"Loud enough?<br/>(Activation Threshold)"}
B -- no --> A
B -- yes --> C{"Long enough to<br/>interrupt?<br/>(Min Speech Duration)"}
C -- yes --> D["Agent stops talking<br/>and listens"]
C -- no --> A
D --> E{"Silence held for<br/>Min Silence Duration?"}
E -- no --> D
E -- yes --> F["Caller's turn ends —<br/>agent answers"]
Min Silence Duration should always be greater than Min Speech Duration. They measure different things — one is a pause, the other is continuous speech — and if speech is the longer of the two, short replies stop working as interruptions: the caller says "no" and the agent keeps going.
The form warns you inline when they are the wrong way round:
Min Speech Duration is longer than Min Silence Duration: short replies such as "yes" or "no" will not interrupt the agent. The presets set silence longer than speech.
Every preset satisfies this. If you are typing values by hand and are not sure, click Conversational and adjust from there.
Each field accepts 0 to 10 seconds, and Activation Threshold accepts above 0 up to 1. A value outside those bounds is not silently honoured — calls fall back to the default, and the form says so:
Min Silence Duration 500 is not a usable value in seconds (0–10); calls use the default 0.55s.
The View Voice Agent (Production) page prints the effective value —
what a call will actually use — rather than the stored number, so a
substitution is visible rather than something you discover on a call:
0.55s (stored 500 is out of range; default applied).
If turn-taking values on an agent were ever corrected for you, the same card notes when, why, and what the previous values were.
If you had an editor tab open when values were corrected, saving from it will not undo the correction: the save lands, and you are told "Turn-taking values were left at their corrected settings; this tab was opened before they were corrected." Reload the page and the fields are editable as normal.
This is the most common tuning request, and the lever is almost always Min Silence Duration. Work through it in this order:
Do not raise or lower Min Speech Duration to fix latency. It governs interruptions only, and changing it will not make the agent answer sooner.
A reply usually starts within a couple of seconds. When the model takes longer, or a tool runs long, the agent does not leave the line silent: it says a short line in its own voice, such as "One moment.", and the reply follows as soon as it is ready. If the reply is ready while the filler is still playing, the agent finishes the filler first rather than talking over itself. That usually takes a fraction of a second. A long custom line is stopped cleanly instead of making the caller wait, and no filler starts once the reply is already on its way.
Two fillers exist, and both are yours to shape on the Voice tab's Fillers card:
| Field | What it does |
|---|---|
| Slow reply: speak after (sec) | How long the caller may wait for the agent's first word before it says a filler. Blank uses the platform default of 3.5 seconds; 0 turns it off. |
| Slow-reply words | Agent decides or My lines: who chooses what the slow-reply filler says. See Who chooses the words below. |
| Slow-reply lines | Up to three short lines, one per row, in the agent's language, rotated so a caller never hears the same one twice running. Blank uses the platform's line. Under Agent decides the field is labelled Example and fallback lines. |
| Use my own wording for another language (optional) | Your own slow-reply lines for a particular language, used instead of the translation. Most agents never need it. |
| Tool wait: speak after (sec) | How long a tool may run in silence before the agent says the tool-wait line. Blank uses each tool's own platform delay. |
| Tool-wait words | Agent decides or My lines, for the line spoken while a tool runs. |
| Tool-wait line | The general line spoken while a tool is still running, with per-tool overrides below it. |
Each filler has its own choice, so you can let the agent write one and keep your own words for the other.
Writing a slow-reply line takes up to about two and a half seconds, so with Slow reply: speak after set below 3.1 seconds the agent always says your lines. A tool that answers within a fraction of a second never asks for a line at all.
A new agent starts on Agent decides for both. An existing agent keeps your lines wherever you wrote some, and moves to Agent decides where the lines were blank. Change either on the Fillers card at any time.
What the platform keeps fixed for a line the agent writes: at most eight words, plain words, and never an answer, a number, a time, a price, a promise or a question. A line that breaks any of these is not spoken, and your line is said instead. The agent writes from the call's language and the last thing each side said; it never sees what a tool was asked to look up.
Write each line once, in the agent's language. When a call is in another language (the caller switched, or Auto-Detect Language heard it), the agent says the line in that language, translated automatically when the call first needs it. It speaks in the voice the call is using for that language, so a language with its own voice under Multi-Lingual Support hears the filler in that voice. Until a translation is ready the agent says nothing rather than speak the wrong language, so the first filler after a switch to a new language may be skipped.
What the platform keeps fixed:
Fillers play on phone calls only.
Every filler is written into the transcript as agent speech, in the words the caller heard. The call's Latency card shows Fillers spoken with how long the caller had waited before each one, marked (written) where the agent wrote the line. Under Agent decides it also shows Filler words: how many lines the agent wrote and how many came from your lines, and why (for example, not ready in time). That lets you read whether each setting is doing its job.
When the business is done and the caller signals they are finished, the agent says goodbye and hangs up. It does not hang up the instant its goodbye ends. It listens for about a second and a half, a little longer if the caller is mid-word.
Each time the agent ends a call, or keeps one open because the caller spoke, the call's transcript shows an end_call entry saying which happened. See Calls, Transcripts and Recordings.
Call handling → Turn-taking & silence → When the line goes quiet has three whole-second fields:
| Field | Default | What it does |
|---|---|---|
| Caller never speaks (sec) | 60 | Hang up if the caller says nothing at all after the greeting |
| Check in after (sec) | 150 | Check in with the caller after this much silence (see below for the words) |
| Hang up after (sec) | 180 | Hang up after this much silence, once the caller had spoken |
Check in after and Hang up after count from the caller's last words, so with the defaults the agent checks in at 2½ minutes and hangs up at 3. Caller never speaks counts from the start of the call (on an outbound call, from the answer). A blank field or 0 uses the default; neither turns the rule off. If Hang up after is not longer than Check in after, the call ends before the check-in is due, and the card's first line says so.
They exist because a caller who puts the phone down without hanging up otherwise holds the line — and the minutes — open indefinitely.
Set them longer for calls where silence is legitimate (someone fetching a document, a caller on hold at their end) and shorter for high-volume lines where an abandoned call is just cost.
These rules are suspended while the caller is meant to be waiting, and the silence counts start again from zero when the wait ends:
By default the agent decides. When Check in after passes it checks in with one short question in its own words and in the language of the call, from what is actually happening on the line. If you would rather it say the same words every time, set When the caller goes quiet, just under the three silence fields, to Exact line, and type the line. Leaving the line blank with Exact line selected uses the platform's "Are you still there?".
On Gemini 3.8 Live the presets and the three turn-taking fields above do not apply — Google's own turn detection runs on the call. Call handling → Turn-taking & silence shows four settings instead, each blank for Google's default: Start of speech and End of speech (High or Low sensitivity), Silence that ends a turn (Google recommends 500–800 ms) and Audio kept before speech (0 clips the first word). Caller can interrupt decides whether the caller's speech cuts the agent off mid-sentence; off, the agent finishes every sentence.
Context compression is always on — it is what lets a live-audio call run past Google's 15-minute limit — and is tuned by Karmaflow; to try an agent without it, contact support@karmaflow.ai. Fillers do not apply: what the caller hears while a tool runs is set per tool under Tools → Tools on live-audio calls (see Tools and Knowledge). Dead-air timeouts and background audio work as on any agent.
Turn-taking is not something to judge from one test call. After a change, let the agent take a handful of real calls, then open Agents → Agent Types → Voice Agents, click ⋯ on the agent's row and choose Performance & insights. Four tiles there measure exactly what you just changed:
| Tile | What it tells you |
|---|---|
| Interruptions / call | How often callers successfully cut the agent off. A flat zero on calls where people plainly tried is the Min Speech Duration problem above. |
| Response latency | How long the agent waited before answering — what a caller experiences as "slow". |
| LLM / TTS first byte | How much of that wait is the model composing, not your audio settings. Compare the two before touching Min Silence Duration again. |
| Tool success | Unrelated to timing, but worth a glance: a failing tool stalls a call more than any setting here. |
The tiles only appear once calls have been taken on the current settings.
Then use Place a test call on the Test & Deploy tab to hear it yourself over a real phone path; the browser test panel is not a fair test of telephony timing.