Most settings here apply to KoreVG pipeline voice (speech-to-text, agent, text-to-speech). Realtime speech-to-speech calls ignore many of them, because the model provider owns turn detection and delivery. Each topic states where it applies. For the pipeline and realtime differences, see Pipeline vs. realtime voice.
Where voice settings live
Set a voice behavior at the narrowest level that fits. A narrower level wins over a wider one.Speak: control what the caller hears
Voice output (VOICE:)
A VOICE: block next to a RESPOND overrides text-to-speech for that reply.
In a FLOW step, write the keys in uppercase (
SSML, INSTRUCTIONS, PLAIN_TEXT). A behavior profile VOICE block accepts the same keys.
Details: Voice configuration, FLOW voice configuration, Behavior profile VOICE.
Spoken-style rules and grounding guards
EXECUTION.voice.instructions fills ## Voice Delivery Instructions in the prompt. It replaces the platform voice formatting rules. Restate “plain speech, no lists” in your text.
false and apply to voice channels only.
Details: Voice instructions and grounding guards.
Realtime delivery tone (coordinator_policy)
For realtime voice agents, EXECUTION.voice.coordinator_policy sets language, tone, pace, brevity, interruption, greeting_style, and two accessibility read-back switches. It shapes delivery. It does not change what the agent says.
Details: Voice coordinator policy.
Barge-in and streaming for one output
BARGE_IN (allow or disallow) decides whether the caller can interrupt one spoken output. STREAMING (true or false) decides whether audio starts before the full reply is ready. Set both at the narrowest scope that fits.
VOICE, its template, its GATHER field, the tool reply policy, the step, the agent’s CONVERSATION.listening.barge_in, then the channel setting. An explicit allow beats a wider disallow.
Where each scope reads the policy:
Placing
BARGE_IN in a behavior profile VOICE block, a channel override, a free-standing ON_START or branch VOICE block, or YAML execution.voice is an error (VOICE_BARGE_IN_ILLEGAL_PLACEMENT).
Streaming follows the same scopes. When the call is in derived turn-taking, or the channel Streaming switch is Off, ordinary output is buffered even if you author STREAMING: true.
Details: Scoped barge-in, Scoped streaming.
Filler phrases
On a voice channel the runtime speaks a short filler (for example “One moment.”) while a tool call, reasoning step, handoff, delegation, extraction, or constraint check runs. Override the text withfiller_<operation> keys in MESSAGES:.
{{ }} placeholders. Add a language suffix for a localized variant.
Details: Filler overrides (voice).
Listen: control how the bot hears the caller
Listening mode, barge-in, and streaming
On KoreVG pipeline voice, the listening mode is fixed for the whole call. Barge-in and streaming are decided per spoken output inside that mode.
The mode resolves in this order: the channel’s
listeningMode, the entry agent’s CONVERSATION.listening.mode, then a derived mode. The derived mode is background when barge-in is allowed, otherwise turn_taking. A handoff, step, reply, or tool cannot change it. A LISTENING key in any VOICE block is an error (VOICE_LISTENING_CALL_LEVEL_ONLY).
Caller silence (on_no_input)
CONVERSATION.listening.on_no_input handles a caller who says nothing.
- Recognized speech, or accepted keypad input, resets the retry count. An ASR partial only restarts the window.
- Input the runtime accepted but has not run is not silence. The window waits.
- The runtime waits one more second after the window runs out before it reprompts, in case a partial is on its way.
- A FLOW step can override the policy with
VOICE: ON_NO_INPUT:. Step messages play after the agent message of the same number.ON_NO_INPUTanywhere else is an error (VOICE_NO_INPUT_ILLEGAL_PLACEMENT). on_exhausted: hookruns theon_no_input_exhaustedhook. The hook can holdCALL,SET,HANDOFF, andCRITICAL, but notRESPOND.- Realtime voice uses static prompts only. OpenAI also honors the channel’s
realtimeNoInputEnabled: false.
voiceNoInput.onExhausted (hangup, transfer, or handoff) applies to every agent and step that leaves on_exhausted unset. An agent, profile, or step on_exhausted replaces it, target included. A project handoff needs a HANDOFF or routing rule from the current agent to the target.
Details: Voice turn-taking, Step-level no-input, on_no_input_exhausted, Project defaults.
Back-channel, hold phrases, and barge-in continuation
Three morelistening blocks tune natural turn-taking.
The gateway can also decide that caller speech over a prompt is a back-channel. The runtime accepts that decision in every listening mode, even with no phrases authored.
Details: Voice turn-taking.
End of speech and turn detection
Tell the recognizer how long to wait after the caller stops speaking.Speech recognition for one field (ASR:)
An ASR: block on a FLOW GATHER field tunes recognition while that one field is asked. The call’s normal settings return afterwards.
The field must be the only field in its step.
ASR: works in FLOW steps only. Anything the runtime cannot apply falls back to the call’s normal settings.
Details: Field-level listening settings.
Keypad input (DTMF:)
A DTMF: block on a FLOW GATHER field sets how the bot collects digits typed on the phone keypad.
Each key resolves on its own, in this order:
SECURE_INPUT COLLECT, field DTMF:, behavior profile, agent CONVERSATION.listening.dtmf, project voiceListening.dtmf, platform default. The platform default is 2 seconds between key presses, # as the finish key, and 1 to 12 digits.
Use SECURE_INPUT, not DTMF:, for secrets.
Details: Field-level keypad settings, Finish key pressed alone, Keypad defaults.
Recognition confidence (sttMinConfidence)
Set sttMinConfidence (0 to 1) on the channel connection or in project runtime config as voice.sttMinConfidence. Pipeline voice discards a recognition scored below it. The channel value wins. 0 at channel level turns the check off.
Details: Project defaults.
Collect sensitive input
Secure input (SECURE_INPUT)
SECURE_INPUT: pauses a FLOW step to collect sensitive values, such as a card number, by keypad or speech. It delivers them straight to one tool. The raw values never reach the model, transcript, session state, logs, recordings, or traces. Speech recognition is suppressed and recording is paused while the caller enters them.
Key points:
- Modes. Step
MODEisdtmf,speech, or both. A spoken answer is read as digits. - Readable fields. A field marked
NOT_SECRET: truewithSAVE_ASis saved for later steps. Only aNOT_SECRETfield can useMODE,RECORD, andSPEECH_AS. - Recording.
RECORD: trueresumes call recording while that field is asked. A step pauses and resumes recording each time it moves between a recorded and an unrecorded field. - Spoken answers.
SPEECH_AS: textkeeps the caller’s words instead of reading them as digits. - Outcome routes.
ON_SUCCESS,ON_FAIL,ON_TIMEOUT, andON_CANCELtake a step name only.
Hold, hang up, and transfer
Caller hold
When the phone network signals a hold on a KoreVG call, the agent stays silent and stops listening. The project’scallerHold setting limits the hold. Set it in Studio under Settings > Runtime Config > Caller hold.
The
on_caller_hold_limit hook can hold CALL, SET, HANDOFF, and CRITICAL, but not RESPOND. Runtime also sets read-only session.callerHold values, for example resumed, durationSeconds, cumulativeSeconds, holdCount, and pendingReply.
Ending the call (COMPLETE)
On a KoreVG voice call every COMPLETE ends the call after its message plays, whether or not it is TERMINAL. Caller speech or keys heard while that message plays are ignored. The message plays once, even if the agent that handed back already said the same line.
A COMPLETE condition’s VOICE block can name SIP headers for the closing BYE:
sip_headers anywhere else gives VOICE_SIP_HEADERS_ILLEGAL_PLACEMENT.
Details: COMPLETE, SIP headers on the closing BYE.
Transfer a call to a human
sipTransferId. At transfer time the runtime uses the authored sipTransferId default, then the project’s voice.sipTransferTargetUri, and otherwise fails with a missing-target error.
Details: Voice transfer settings, Custom voice transfer, Genesys SIP transfer headers, Trusted SIP transfer destinations, Inbound SIP header placeholders, Dynamic destinations, RESOLVE_SIP_DESTINATION.
Runtime support and upgrades
An agent that uses a voice setting is published with an IR feature marker. A runtime that does not support the marker refuses to publish, deploy, or run the agent. Upgrade every runtime before you publish such an agent.
A Studio import into a project whose runtime lacks
voice_listening_asr drops the speech block and warns (LISTENING_KEY_NEEDS_NEWER_RUNTIME).
Voice diagnostics
Each code links to its entry. Errors block compile or publish. Warnings do not. Severity matches the entry on the errors page.Delivery, listening, and placement
Silence and hold
Field listening and keypad
Secure input
SIP and transfer
Related
- Voice channels — create and manage voice channels in Studio.
- Behavior profiles —
VOICEandCONVERSATION.listeningin full. - FLOW — field and step voice settings,
SECURE_INPUT, and SIP destinations. - Lifecycle and hooks — no-input and caller hold hooks, filler.
- Errors and warnings — every diagnostic entry.