Voice
The Voice page configures Realtime Dialog — a spoken, voice-to-voice conversation mode that runs alongside your agent's existing text chat rather than replacing it. Users can talk to the AI assistant out loud and hear its responses spoken back, and switch between voice and text at any point. Everything the agent already does — knowledge base lookups, web search, custom tools, and so on — keeps working the same way behind the scenes; only how the conversation is carried in and out changes.
Navigate to Conversation → Voice in the left sidebar.
Realtime voice consumes tokens differently from text chat (audio in, audio out, plus transcription), so enabling it affects the cost of running the agent.
Realtime Dialog
Enable Realtime Audio
Turns Realtime Dialog on or off for this agent. When enabled, end users get a way to start a spoken conversation from the chat interface — via the + menu next to the chat input — in addition to typing as usual.
When this is off, none of the other settings on this page apply and no voice entry point is shown to users.
Default Voice Display
Controls how a voice session is presented the moment it starts:
| Value | What the user sees |
|---|---|
| Immersive | A full-screen, voice-only view: a sound-wave animation reacting to who is speaking, no visible transcript text. Charts, tables, and cited sources produced during the call appear as swipeable pages in a canvas pager instead. |
| Transcript | A text-style view: the spoken conversation is shown as a scrolling transcript of chat bubbles, same as a regular chat thread, with clickable citations and a text box to type into mid-call. |
This is only the starting point — the end user can still switch to the other display from their own preferences once a call is underway.
Immersive View Welcome Text
Only shown when Enable Realtime Audio is on. Free text displayed in the immersive view while the canvas is still empty — before any chart, table, or source page has arrived. Leave blank to use the built-in default placeholder ("Speak to the assistant. Charts, tables and sources will appear here as pages you can browse.").
Canvas Display
Only shown when Enable Realtime Audio is on. During a voice call, some content is better shown on screen than spoken aloud — charts, tables, long lists, code, math, scripture, or anything that reads awkwardly out loud. These settings control when the AI assistant pushes content to the canvas instead of speaking it directly.
Automatic Canvas Content Detection
When enabled, the AI decides for itself, response by response, which parts to speak conversationally and which parts to show on the canvas instead — not limited to charts and tables. Leave this off if you'd rather rely only on the explicit list below.
Mandatory Canvas Content Types
A list of free-text content types (for example chart, table, arabic scripture, mathematical formula, cards) that must always be rendered on the canvas instead of spoken as plain text, whenever a response contains that kind of content. This applies regardless of whether Automatic Canvas Content Detection is on.
Click Add Content Type to add a row, and the delete icon on a row to remove it. This only changes where the content is rendered — the agent still speaks whatever explanation the answer calls for alongside it.
Voice & Transcription
Only shown when Enable Realtime Audio is on. Choose the voice the AI assistant speaks with, and the model used to transcribe what the user says into text.
Voice Persona
Selects the voice used for the assistant's spoken responses. Available options: alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin, cedar.
Click the ▶ button next to the dropdown to hear a sample in the selected voice before saving. The sample speaks your agent's own welcome message, so you hear the voice saying what users will actually be greeted with. If no welcome message has been written yet, a short built-in English sentence is spoken instead.
Transcription Model
Selects the model used to transcribe the user's spoken input into text.
| Model | Notes |
|---|---|
gpt-4o-transcribe | Default. High-accuracy transcription, billed per token. |
gpt-4o-mini-transcribe | Faster and cheaper than gpt-4o-transcribe, also billed per token. |
gpt-transcribe | Billed per minute of audio rather than per token. |
whisper-1 | Legacy transcription model. Kept for agents already configured with it. |
Voice Activity Detection (VAD)
Only shown when Enable Realtime Audio is on. VAD is what tells the system when the user is speaking and when they've finished, so it knows when it's the AI's turn to respond. Without good VAD, the assistant may cut users off mid-thought, or respond to background noise and filler sounds like "hmm" as if they were a real turn.
Smart turn-taking (semantic VAD)
Off by default. When enabled, turn detection is model-based rather than purely energy-based — it notices natural pauses in speech instead of waiting a fixed amount of time, which is better at telling real speech apart from filler sounds and background noise.
Stop assistant on interruption
Off by default. When on, the assistant stops speaking the instant the user starts talking — responsive, but it may cut off short replies like "ok" or "mhm". When off, it waits about a second before stopping.
Sensitivity
When Smart turn-taking is on, this is a dropdown controlling how eagerly the model ends the user's turn: low, medium, high, or auto (default). Higher responds faster but may interrupt more.
When Smart turn-taking is off, this is a slider from 0 to 1 (default 0.7) setting the energy-based activation threshold. Higher ignores more background noise; lower responds faster, at the cost of more false interruptions.
Push-to-talk
Only shown when Enable Realtime Audio is on. Push-to-talk requires users to hold down a button while they speak — the microphone only streams audio while it's held. This guarantees background noise or other conversations can never be mistaken for a real turn, which is especially useful in noisy environments or when extra privacy is needed.
Require hold-to-talk
Enabled by default. Shows a hold-to-talk button in the voice call UI. When enabled, the user can toggle at runtime between this and automatic (VAD-driven) listening; when disabled, the call always runs in automatic mode.
Start calls in
Only shown when Require hold-to-talk is enabled. Which mode a voice session opens in (default Automatic):
| Value | Behavior |
|---|---|
| Automatic | The server listens continuously and detects turns using the VAD settings above. |
| Push-to-talk | The user must hold the talk button to be heard; nothing is streamed otherwise. |
Saving Changes
Click Save after adjusting any setting. Publish your agent to make the changes visible to end users.
Related pages
- Chat — language and formatting settings for text chat
- Behavior — mission and personality, which also shape how the agent speaks during a voice call
- Model & Context — AI model selection and conversation context memory