Talk by Voice

Open in CMS

In this step you let people talk to your agent out loud and hear it answer. Voice runs alongside your agent's existing text chat rather than replacing it: users can switch between speaking and typing at any point, and everything the agent already does (knowledge base lookups, web search, custom tools, other agents) keeps working the same way behind the scenes. Only how the conversation is carried in and out changes.

In the CMS, voice is configured on the Agent Voice page: open Conversation β†’ Voice in the left sidebar. It is one scrolling page of sections, saved together with one Save all changes. At the top, the Voice conversations section holds the switch and, while voice is on, a sticky bar of chips, one per section; click a chip to jump to its section. On a phone the chips are a menu that also shows the current choice of each section (style, where the voice runs, voice, screen, start screen).

SectionWhat it holds
Conversation style and Call flowThe conversation style, and right below it the call flow of that style: turn-taking and push-to-talk for Turn-by-turn, or talk-over, listening sounds and call limits for Natural conversation
ModelsVoice runs on, and the Answering AI for Natural conversation
VoiceThe assistant's voice and speaking style
Call screenThe voice screen style and everything the immersive view shows
On-screen content and Sending what is on screenWhat goes on the canvas, and sending it on
Start screen and LinksThe start screen and the chat and voice links
Agent voice page for Atlas: the Voice conversations section with the switch Let users talk to the agent On, and under it the bar of chips Style, Call flow, Models, Voice, Call screen, On-screen content and Start & links

In this section you give Atlas, Kirana Group's operations assistant, a voice. A site supervisor with gloves on can then ask "Who signs the permit for work above two metres?" out loud and hear the answer, while approval chains and project status appear on screen. Atlas uses Turn-by-turn, which keeps quiet calls cheap; Natural conversation is the style marked Default. Natural Conversation Settings shows what changes when you switch to Natural conversation.

Voice calls use credits differently from text chat, and differently for each conversation style: Turn-by-turn uses credits for the audio and text it hears and speaks (plus transcription), while Natural conversation uses credits for every minute the call is open (silence included) plus the answers its Answering AI writes. Either way, enabling voice affects the cost of running the agent. See Billing.


Before you start


Turn on voice

Voice conversations

The switch row is called Let users talk to the agent. It turns voice on or off for this agent. When on, 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. They can also open a call straight from a link ending in /voice, and you can make voice the screen your agent's link opens on: see Start screen.

When this is off, none of the other settings on the page apply and no voice entry point is shown to users. The other sections appear only once this switch is on.

Switching Voice conversations on also selects Natural conversation under Conversation style where it can run: when the style is offered and Natural conversation isn't blocked for your organization or agent. Otherwise the style stays Turn-by-turn. An agent already saved with Natural conversation keeps it. Nothing is saved yet, so you can still pick Turn-by-turn before you click Save. Clark, the configuration assistant, does the same when it switches voice on for you. Keep the cost in mind: Natural conversation on Qlar uses credits for every minute a call is open (see Billing); on your computer or your own account no Qlar credits are charged per minute.

A note under the conversation style points to Conversation Initiative on the Behavior page. That setting decides whether the agent offers a next step after answering, in calls and chat alike. In a call the offer is spoken: one short offer, never a list.

The Voice screen style, which decides what users see during a call, is in the Call screen section. It is explained in Voice Screen and On-Screen Content.


Conversation style

The first section after Voice conversations. Only shown when voice is on and Natural conversation is available to your organization (from Qlar, or on your computer) or the agent already uses it. Chooses how the assistant talks with users in voice calls:

OptionWhat it is
Natural conversation (listed first, marked Default)Like a phone call with a person: it listens while it speaks. It makes short listening sounds ("mm-hmm") while the user talks, notices when it is talked over and yields, and keeps the conversation flowing. Anything that needs your knowledge base, data, tools, or the screen is handed to an Answering AI model, and the voice tells the user the result in its own words.
Turn-by-turnOne person talks at a time: the assistant waits until the user has finished speaking, then answers. How it recognizes the end of a turn is tuned in Conversation flow and Push-to-talk. Marked Default instead where Natural conversation can't run.

Right below the style, the Call flow section shows the settings of the style you picked, labelled with its name. Switch the style and the section changes with it, so you see at once what the choice means. Each style only reads its own call-flow settings; the other style's stay saved.

Conversation style section for Atlas: the option cards Natural conversation (Default) and Turn-by-turn, selected, each with a short description and how it uses credits, the note about follow-up offers on the Behavior page, and below it the Call flow section labelled Turn-by-turn with Speech recognition gpt-transcribe, Detect when the user has finished, Noise filter and the closed fold Advanced

Natural conversation is available to selected organizations; the option only appears where it is available. It is not offered to organizations with Privacy Mode on, because during a Natural conversation call the conversation β€” what the user says or types, attached images, and the answers β€” can be kept by the model provider, where privacy mode can't remove it.

If Natural conversation is later switched off for your organization, an agent that already uses it still shows the choice with a warning: its calls run Turn-by-turn until Natural conversation is available again, and its Natural conversation settings are kept. If its voice runs on your ChatGPT subscription, calls end instead until voice on your subscription is available again; they never switch to Qlar voice or Turn-by-turn.

If Natural conversation isn't available to your organization, the section is titled Voice model and the page offers only Turn-by-turn's Voice runs on. Everything on this page and in Voice, Listening and Turn-Taking still applies.

Voice runs on

Each conversation style has its own Voice runs on, in the Models section for the style you picked. Beside the label, an info mark explains the three choices:

ChoiceTurn-by-turnNatural conversation
Organization default (where every agent starts)What your organization picked on Organization β†’ AI Models β†’ Tasks under Voice calls Β· Turn-by-turn; Qlar voice until it picks oneWhat it picked under Voice calls Β· Natural conversation; Qlar voice (GPT-Live) until it picks one
Qlar voiceQlar voice (gpt-realtime-2.1) or the lower-cost Qlar voice mini (gpt-realtime-2.1-mini)Qlar voice (GPT-Live), whatever the organization picks
Your own accountA realtime model your organization added under its own OpenAI or Azure OpenAI keyA GPT-Live model your organization added under its own OpenAI or Azure OpenAI key
Your computerβ€”A connected computer, through Codex and the ChatGPT subscription signed in on it; see below

On your organization's own account no Qlar credits are charged for the voice. See Voice, Plugins and Services on Your Own Models. While your organization's rules allow only its own AI accounts, Qlar's choices are greyed out.

How the two styles differ

Turn-by-turnNatural conversation
ConversationOne person talks at a timeListens while speaking, makes listening sounds, yields when interrupted
Who answersThe voice itself, using the agent's toolsThe Answering AI (by default the chat AI); the voice speaks its answer
Voices10, default marin22, default cedar (see Assistant's voice)
Speech recognitionThe Speech recognition model you choose, which uses credits separatelyBuilt in, no credits of its own
Turn-taking settingsConversation flowWhen someone talks over the assistant and Listening sounds
Push-to-talkAvailableNot available: the call always listens
Images during a callAlways readDepend on the Answering AI (see Attachments)
Call limitsNoneEnds after a quiet period and at a maximum length (see Call limits)
BillingCredits for the audio and text heard and spoken, plus transcriptionAbout 1,500 credits per minute while the call is open, silence included, plus the Answering AI's credits (see Billing)

Good to know when switching:

  • Each style keeps its own settings. The Turn-by-turn settings stay saved while Natural conversation is selected, so switching back loses nothing.
  • Save, then Publish. The CMS preview uses the new style as soon as you save; users get it once you publish. A call that is already running keeps the style it started with.
  • Everything else applies to both: the voice screen style, on-screen content, and speaking style work the same way in either style. Attachments and immersive view instructions differ slightly; their sections explain how.
  • Turn-by-turn is the fallback. If Natural conversation can't take a call (for example while it is at capacity), the call runs Turn-by-turn with your Turn-by-turn settings, so users can still talk. The exception is voice on your ChatGPT subscription: those calls end instead.

Which one should you pick? Natural conversation is the default: calls feel like talking to a person, and every minute a call is open uses credits (no Qlar credits per minute on your own account or your computer). Pick Turn-by-turn when a quiet call should cost little or users call from noisy places and need push-to-talk. Switching voice on selects Natural conversation where it can run, so click Turn-by-turn before you save if that is the one you want.


Billing

Turn-by-turn uses credits for what it actually processes: the audio and text it hears and speaks, plus the transcription of what the user says. A call where nobody talks costs little.

Natural conversation uses credits in two parts:

PartRateWhat counts
Voice timeAbout 1,500 credits per minute (25 per second)Every second the call is open. Silence, time the user is muted, and time the assistant has been asked to stay quiet all count.
AnswersThe Answering AI's normal token rateCounted exactly like a chat answer from that AI, each time the voice hands something over; twice the rate for a + model, which runs on Fast mode. Tools the answer uses (web search, knowledge base, plugins) use credits as usual.

For example, a 5-minute call in which the voice hands four questions to Smart costs about 7,900 credits: 7,500 for voice time and roughly 100 per answer (the exact amount depends on how long the answers and the conversation are). A question that needs a lookup costs a little more, because the tool's result is sent back to the model, and tools such as web search or the knowledge base add their own small charges. On the usage pages, voice time appears under the category dialog-live and the answers under dialog-live-backend.

The call limits keep an idle or forgotten call from running up voice time.


Voice on your ChatGPT subscription

If your organization has connected your ChatGPT subscription through Codex on one of its computers, a Natural conversation call can run on the realtime voice of the ChatGPT account signed in on that computer instead of Qlar's GPT-Live. Your own ChatGPT subscription, or an API key used by Codex, pays for the voice, and no Qlar credits are charged per minute.

In the CMS this choice is called Your computer. After choosing Natural conversation in the Conversation style section, pick where it runs under Voice runs on in the Models section: Organization default, Qlar voice, Your own account (a GPT-Live model your organization added under its own key; see Natural conversation on your OpenAI account), or Your computer: name for one of the organization's connected computers, with its online state. A computer that can't take calls is greyed out with the reason β€” offline, not connected yet, its tools can't do voice, or Codex isn't signed in.

On your subscription
Assistant's voiceThe voices offered by the account on that computer, starting with its default, marked "(subscription default)". There is no preview.
Calls at a time2 at once per connected computer by default (the Voice page shows its own limit); another caller sees that voice calls are busy.
Computer offline or removedThe call ends. It never switches to Qlar's voice or to Turn-by-turn, so there are no surprise credits.
ChatGPT subscriptionNeeds outbound UDP port 3478 from that computer.
BillingNo Qlar credits per minute. The Answering AI uses credits as usual, and none when it is a subscription model.

Voice on your subscription is not available while Privacy Mode is on: an agent whose voice runs on your subscription has its calls end until Privacy Mode is off or you pick Turn-by-turn. Everything else on this page β€” the voice screen, on-screen content, speaking style and the Natural conversation settings β€” works the same way.


Saving changes

Click Save all changes after adjusting any setting, then confirm. It saves every section at once, and the bar says Unsaved changes until you do; the CMS then says Voice settings saved. The CMS preview uses your saved settings right away; publish your agent to make the changes visible to end users. A section whose settings need attention (for example an Answering AI that is no longer available) shows a notice.


In this section

PageWhat you configure
Voice, Listening and Turn-TakingThe Voice section (voice and speaking style), and Turn-by-turn's Call flow section: speech recognition, turn-taking and push-to-talk.
Voice Screen and On-Screen ContentThe Call screen, On-screen content and Start screen sections: immersive or transcript view, attachments, what goes on the canvas, instructions for the immersive view, subtitles, where the progress indicator appears, the opening sequence of a new call, and the start screen.
Natural Conversation SettingsNatural conversation only: the Answering AI (Models section), and its Call flow section: how the assistant yields when talked over, listening sounds, and call limits.

Start with Voice, Listening and Turn-Taking to pick a voice, then set up the voice screen. Open Natural Conversation Settings only if you chose Natural conversation above.


Check the result

  1. Turn on Voice conversations. If Natural conversation gets selected, click Turn-by-turn in the Conversation style section to keep Atlas on it, and check that Call flow now says Turn-by-turn. Then click Save all changes and confirm.
  2. Open the preview of your agent, click the + menu next to the chat input, and start a voice conversation.
  3. Ask a question your agent can already answer in text. For Atlas, say "Who has to approve a project worth $2.5 million?". You should hear the approval chain from SOP-FIN-007 (Group Risk Office review, then Regional Director, then CFO) in the voice you chose, and see the call in the voice screen style you set.
  4. When you're happy, publish the agent so users get voice too.

Next steps