Natural Conversation Settings
Open in CMSOn this page you tune voice calls that use Natural conversation: which AI answers the questions, how readily the assistant yields when it is talked over, how often it makes listening sounds, and when a call ends on its own.
These settings are on the Agent voice page (Conversation β Voice in the CMS sidebar), which is one page of sections: the Answering AI in the Models section, and the rest in the Call flow section, right below Conversation style. The chips at the top of the page jump to each section.
Natural conversation only. The Call flow section shows these when the conversation style is Natural conversation, in place of Turn-by-turn's Conversation flow and Push-to-talk.
Before you start
- Turn on Voice conversations and select Natural conversation as the conversation style. It is only offered to selected organizations without Privacy Mode.
- Read how Natural conversation uses credits: every minute a call is open counts, silence included.
- Check which Conversation AI model your agent uses; it is the default Answering AI.
Switch to Natural conversation
In this example Atlas, Kirana Group's operations assistant, runs its calls Turn-by-turn. This page shows what changes when you switch it to Natural conversation.
In the Conversation style section, select Natural conversation. Right below it, the Call flow section switches to Natural conversation: speech recognition, turn-taking and push-to-talk make way for talk-over, listening sounds and call limits. Atlas's Turn-by-turn settings stay saved, so switching back later loses nothing; on Qlar they are also what a call uses if it ever falls back to Turn-by-turn.

Answering AI
In Natural conversation the voice carries the conversation itself (greeting, listening, small talk, clarifying questions) and hands everything else to the Answering AI: questions that need your knowledge base, data, or tools, and anything shown on the canvas. That AI works out the answer, and the voice tells the user in its own words.
It is set in the Models section, below Voice runs on, once Natural conversation is selected. Voice runs on chooses where the voice itself runs: Qlar voice (GPT-Live on Qlar's account), Your own account (your Azure OpenAI or OpenAI account), or Your computer (a connected computer, through Codex). The info mark next to the label explains the three.

For Atlas, the defaults are a good start: the Answering AI follows the chat AI (Smart), Balanced yields to real interruptions, Moderate listening sounds, and calls end after 2 minutes of silence or 30 minutes in total.
| Choice | Behavior |
|---|---|
| Same as chat AI (β β¦) (default) | Uses the agent's Conversation AI model; the arrow shows which model that is today. Change the chat AI later and calls follow. This is the agent's own model, not one an end user picked in the chat's model switcher. |
| Any other model | Any model available to your organization, including your organization's custom models. Only calls use it; text chat keeps the Conversation AI model. |
Below the field, the page describes the selected model, shows its estimated credit ratio, and lists what it can take in during a call as chips, such as Image in and Document in.
- Thinking depth comes from the model. There is no separate setting. Light answers without reasoning first and is the fastest; Smart and Analyst reason briefly; Insight, Thinker, and Sage reason longer: better at hard questions, but with longer pauses β with Thinker and Sage, the page notes that callers may wait several seconds.
- + models are the same model on Fast mode. Light+, Smart+, Analyst+, Insight+, Thinker+, and Sage+ think exactly like the model without the +, but answer about 1.5β2x faster, which shortens the pauses in a call. Their answers cost twice the token rate. Under heavy load an answer can occasionally run at normal speed; it is then billed at the normal rate.
- Custom models run through Qlar. Qlar's built-in models are run by OpenAI right next to the voice, the fastest path. Custom models, and models from other providers, run through Qlar exactly as they do in chat, so their answers may start a moment later; the page says so when you pick a custom model.
- Credits like chat answers. Each answer uses credits at the chosen model's normal token rate, as a chat answer from that model would, plus any tools it uses, on top of the call minutes. See Billing.
- If the model becomes unavailable β for example, a custom model is disabled β calls use the chat AI instead. The field then shows the stored model as (unavailable, calls use the chat AI), with a note naming the model calls use meanwhile.
The Answering AI also decides whether images can be attached during a call; see Attachments during an immersive call.
When someone talks over the assistant
How readily the assistant stops speaking when the user starts talking.
| Value | Behavior |
|---|---|
| Responsive | Stops talking as soon as someone starts speaking. |
| Balanced (default) | Stops for a real interruption and keeps listening while the user pauses to think. |
| Patient | Stops only for a clear new request; ignores short sounds like "ok" and background noise. Suits noisy places. |
Listening sounds
Short sounds such as "mm-hmm" that the assistant makes while the user talks, to show it is following. This sets a style, not an exact frequency.
| Value | Behavior |
|---|---|
| Off | Stays quiet while the user is speaking. |
| Light | Only during longer turns. |
| Moderate (default) | Now and then, without competing with the user. |
| Frequent | Often, to show it is following. |
Call limits
Because Natural conversation uses credits every second a call is open, even when nobody speaks, every Natural conversation call has two limits. Neither can be turned off. Open the Advanced fold of the Call flow section to see them.
| Setting | Choices | What happens |
|---|---|---|
| End the call after silence | 1, 2 (default), 5, 10 or 15 minutes | Once nothing has happened for this long β nobody speaking, nothing typed or attached, no answer being prepared β the assistant asks whether the user is still there. If there is still no reply about 20 seconds later, it says goodbye and ends the call. The user sees Call ended after a quiet period and can reconnect. |
| Maximum call length | 10, 20, 30 (default), 60 or 120 minutes | The assistant warns about a minute before the limit, then ends the call; the user sees Call time limit reached. Reconnecting doesn't reset the clock: the time counts across reconnects of the same call, and reconnecting to the same conversation within about 10 minutes after the limit ends again right away. The user can start a new conversation instead. |
A value set outside these choices through the API (anything from 30 seconds to 15 minutes, and from 1 to 120 minutes) is kept and shown as an extra choice.
Below the limits, the page shows the most voice time one call can cost: Per call up to about 45,000 credits of voice time. The answering AI uses credits too. for the default 30 minutes (the maximum length Γ 1,500 credits).
Check the result
- Click Save all changes. The CMS preview uses Natural conversation right away; users get it after you publish.
- Start a voice conversation in the preview (+ menu next to the chat input).
- Ask something that needs your knowledge base or data. For Atlas, say "Who has to approve a project worth $2.5 million?". The voice hands the question to the Answering AI model, then tells you the answer in its own words: Group Risk Office review, then Regional Director, then CFO, from SOP-FIN-007.
- Talk over the assistant while it speaks and check that it yields the way you chose. Listen for the listening sounds while you talk.
- To go back, select Turn-by-turn in the Conversation style section and click Save all changes. Your Turn-by-turn settings are still there.
Next steps
- Next in the journey: Remember Each User.
- Voice, Listening and Turn-Taking: choose one of the 22 Natural conversation voices and the speaking style.
- How Credits Work: how voice calls show up in your credit usage.