The Agent Voice screen lets users talk to your agent and hear it talk back. Found in the agent sidebar under Conversation → Voice. Every card below the first appears only while Voice conversations is on.
Turn on voice
| |
|---|
| Voice conversations | Lets users speak instead of typing; text chat still works too. Off by default. Every other setting on this page appears only when this is on. |
| Conversation Initiative note | Whether the assistant offers a next step after answering is set by Conversation Initiative on the Behavior page, for calls and chat alike. |
| Voice screen style | Starting display for a new voice call: Immersive (animated view with sound waves and cards, canvas pager) or Transcript (plain text log, like a chat). Users can change it anytime during a call. |
| Start screen | What the agent's own link opens: Text chat (default) or Voice, straight into a call. Website widgets, the CMS preview, WhatsApp Redirect to web and notification links stay on text chat. Below it, the Chat link (…/chat) and Voice link (…/voice) with a copy button open that screen whatever the start screen is. See Start screen. |
| Welcome message | Shown in the immersive view before anything else appears. Leave blank for a built-in line telling callers to speak and that charts and sources will appear there. |
Conversation style
Shown only when Natural conversation is available to your organization (from Qlar or on its ChatGPT subscription), or when the agent already uses it.
| |
|---|
| Turn-by-turn | Default. One person talks at a time: the assistant waits until the user finishes, then answers. Tuned by Conversation flow and Push-to-talk. Billed by the audio and text actually spoken and heard (plus transcription), so a quiet call costs little. |
| Natural conversation (Beta) | Like a phone call: keeps listening while it talks, makes short listening sounds, and passes questions that need your knowledge, tools or the screen to the Answering AI model. Billed per second while the call is open, silence included (about 1,500 credits per minute), plus the Answering AI model's usage like a chat answer, plus tools. |
| Availability | Not available while the organization's Privacy Mode is on, when it is not enabled for the organization, or while it is switched off on the platform. If an agent set to Natural conversation cannot use it, a warning says why and calls use Turn-by-turn until it is available again; its settings are kept. An agent whose voice runs on your ChatGPT subscription is the exception: its calls end until voice on your subscription is available again or you pick Qlar or Turn-by-turn (see Voice runs on). |
| Switching | Each style keeps its own settings, so switching back loses nothing. The CMS preview uses the new style as soon as you save; users get it after you publish. Calls already in progress keep their style. |
Voice runs on
Shown under Conversation style when Natural conversation is selected and the organization has connected its ChatGPT subscription on a computer and may use it for voice (or the agent is already set to one). Computers that can't take a call right now are listed but disabled. See Use Your Claude or ChatGPT Subscription.
| |
|---|
| Qlar | Qlar's GPT-Live, with its voices and per-minute credits. |
| A connected computer | One option per connected computer, with ● online or ○ offline and how it signs in, for example "Uses the ChatGPT subscription signed in on that computer · no Qlar credits per minute". Disabled with the reason when it can't take calls: offline, not connected yet, its tools can't do voice, or its Codex isn't signed in. |
| Not available | A line under the options says why when voice on your subscription is off: switched off on the platform, subscriptions turned off for the organization by Qlar, or privacy mode on. |
| Warning | For a connected computer: how many calls it takes at once (2 by default; another caller sees that voice calls are busy), that calls end while it is offline and never switch to Qlar voice, that a ChatGPT subscription needs outbound UDP port 3478, and that there are no Qlar credits per minute. |
| Assistant's voice | On a connected computer, lists the voices offered by its account, starting with "(subscription default)". No preview. |
On-screen content
| |
|---|
| Let the assistant decide | When on, the assistant decides each time it replies what to say out loud and what to show on screen (the canvas). |
| Always show on screen (never spoken) | Content types that always appear on the canvas, even when Let the assistant decide is on. Pick from chart, table, arabic scripture, mathematical formula and cards, or type your own. |
| Voice calls only | These settings apply to voice calls only; the Charts and Table settings on the Chat page do not apply here. |
Sending what's on screen
During a call, a signed-in user can ask the assistant to send what is on the canvas to someone. The assistant reads the number or spells the address back before sending, and the recipient sees the user's name. Users who are not signed in are asked to sign in first. Pictures arrive as images; charts and tables arrive as a link that lasts 7 days. Sending is turned off while the organization's Privacy Mode is on.
| |
|---|
| Email | Not a setting: always available. The user names an address, or says "to my email" for the address of their account. It comes from the agent's own {agent}-noreply@qlar.ai address. |
| Send to WhatsApp | Off by default. When on, the user can also have it sent to a WhatsApp number, and it arrives from this agent's own WhatsApp number. Needs a WhatsApp channel on the public server or a private server; Meta Cloud, Coster and Qontak can only start a conversation with an approved message template and are not supported yet. Without such a channel the switch is locked, with a link to Channels → WhatsApp. |
| Which number it comes from | Users' calls send from the Published configuration's WhatsApp channel. Preview calls send from the Draft configuration's channel, or from the Published one when the draft has none. A warning appears when only the Draft configuration has a channel that can send. |
| Limits | The channel's contact rules apply (blacklist, private mode). At most 10 per recipient and 30 in total per hour from one channel. |
| Deleting the channel | Deleting a WhatsApp channel while Send to WhatsApp is on shows a warning first. Deleting the agent's last channel that can send also turns Send to WhatsApp off. |
Immersive view instructions
| |
|---|
| What the assistant should do in the immersive view | Free text added to the assistant's instructions only for calls shown in the immersive view, never for a transcript-view call, since there the answer is already readable as text. Written as instructions to the assistant, not as text users see (e.g. "always show database figures as a chart"). Applied after the On-screen content settings above, so it wins where the two disagree. Blank adds nothing. In Natural conversation it also guides the Answering AI model; the voice reads only the first 4,000 characters, and a warning appears when the text is longer. |
| Transcript note | When Voice screen style is Transcript, a note explains that these instructions only apply to users who switch to the immersive view themselves. |
Subtitles
Immersive view only; the transcript view always shows the conversation as text. Subtitles appear on a dark glass panel over the call buttons while someone speaks, fade out a moment after, and hide when tapped until the next exchange.
| |
|---|
| User transcription | Off by default. When on, the user's own words appear in light grey as they speak. |
| Assistant subtitles | Off by default. When on, the assistant's words appear in white, in step with its voice; an interrupted answer stops at the last word heard. |
| Transcript note | When Voice screen style is Transcript, a note explains that subtitles only appear for users who switch to the immersive view themselves. |
Progress indicator
Immersive view only. Where the immersive view shows what the assistant is working on while it looks something up or prepares a card.
| |
|---|
| Top left | Default. A line under the assistant's name in the top bar, with a spinner beside the name. |
| Center of the screen | A rounded, translucent box over the middle of the screen, in the opposite colours of the call screen, with a spinner and a line whose text shimmers from left to right. The top bar then shows no progress. The box fades away a moment after the assistant finishes. |
| Transcript note | When Voice screen style is Transcript, a note explains that this only applies for users who switch to the immersive view themselves. |
Opening sequence
Immersive view only, and only for a new call (not one that reconnects). The call opens with a short sound, then the assistant reads the introduction aloud and shows the example messages one by one, each as it is said. A real question or remark from the caller, or a tapped example, stops it at once.
| |
|---|
| Play the opening sequence | Off by default. When on, new immersive calls open with the sequence. |
| Play opening sound | On by default. Off skips the music: the assistant starts speaking after a short pause. |
| Opening sound | One of ten short sounds (about four seconds), default Cinematic Rise. The play button beside it plays the sound picked. |
| Read the introduction aloud | On by default. Off skips the introduction message: the assistant goes straight to the lead-in and the example messages. |
| Source of the words | The introduction message and example messages on the Introduction page, with variables filled in and buttons left out. A warning appears when there is nothing to read. |
| Transcript note | When Voice screen style is Transcript, a note explains that this only plays for users who switch to the immersive view themselves. |
Voice and listening
| |
|---|
| Assistant's voice | The voice the assistant speaks with. Turn-by-turn: 10 voices, default marin; the play button speaks the agent's welcome message, or a built-in sample sentence when that is empty. Natural conversation: 22 voices, default cedar, some with a language and accent note; the play button plays a fixed sample of the voice. |
| Voice for voice-note replies | Natural conversation only: the Turn-by-turn voice list, used when the assistant answers with a voice note (e.g. WhatsApp). Voices offered only in Natural conversation can't be used here. |
| Speech recognition | Turn-by-turn only. The model that turns what users say into text: gpt-transcribe (default), gpt-4o-mini-transcribe, gpt-4o-transcribe or whisper-1. Natural conversation understands speech by itself, not billed separately. |
| Speaking style | Everyday (default; conversational, with an occasional "hmm" or filler), Expressive (livelier and warmer) or Formal (polite and businesslike, no fillers or slang). Applies to both conversation styles. |
Natural conversation settings
Shown only when the conversation style is Natural conversation.
| |
|---|
| Answering AI model | The model that answers anything needing your knowledge base, tools or careful reasoning; the voice speaks its answer. Default: Same as chat model (the Conversation AI Model on Model & Context). Other organization models can be picked, including custom ones; a custom model runs through Qlar and answers may start a moment later. Deep-thinking models (Thinker, Sage) show a warning that callers may wait several seconds. If the picked model is no longer available, calls use the chat model. Billed like a chat answer, on top of the call minutes. |
| When someone talks over the assistant | Responsive (stops as soon as someone speaks), Balanced (default; stops for a real interruption and keeps listening while the user pauses to think) or Patient (stops only for a clear new request, ignoring short sounds like "ok" and background noise). |
| Listening sounds | Short "mm-hmm" sounds while the user talks: Off, Light, Moderate (default) or Frequent. A style, not an exact frequency. |
| End the call after silence | 1, 2 (default), 5, 10 or 15 minutes. When nobody speaks, types or uses the screen for this long, the assistant checks in, then hangs up. Cannot be turned off. |
| Maximum call length | 10, 20, 30 (default), 60 or 120 minutes, counted across reconnects; the assistant wraps up a minute before the end. Cannot be turned off. |
| Per-call cost estimate | Shows the most voice time one call can cost at the chosen maximum length, plus AI model usage. |
Conversation flow
Turn-by-turn only; hidden for Natural conversation, which has no turn detection to tune.
| |
|---|
| Smart turn-taking | Notices natural pauses in speech (semantic VAD) instead of waiting a fixed amount of time. Off by default. |
| Stop assistant on interruption | On: the assistant stops instantly when the user talks, but may cut off short replies like "ok". Off (default): it waits about a second before stopping. |
| Sensitivity (smart turn-taking on) | Auto, Low, Medium or High. Higher responds faster, but may interrupt more. |
| Sensitivity (smart turn-taking off) | Slider from 0 to 1, default 0.9. Right ignores noise; left responds faster. |
Push-to-talk
Turn-by-turn only; a Natural conversation call always listens and never shows a hold-to-talk button.
| |
|---|
| Require hold-to-talk | When on, users hold a button while they speak, so background noise is never picked up; users can switch between this and automatic listening during the call. Off by default. Turning it on sets Start calls in to Hold-to-talk. |
| Start calls in | Only shown when hold-to-talk is on. Which mode a call opens in: Automatic or Hold-to-talk. Users can switch anytime during the call. |
Saving
| |
|---|
| Save | Sticky button at the bottom of the page. Saves all cards at once. |
Attachments During an Immersive Call
Not a card on this page: attachments use the settings on Model & Context.
| |
|---|
| Attach button | Right of the microphone in the immersive view. Picked images/documents preview above the controls and are sent with the user's next spoken turn (or push-to-talk release); removable until then. |
| Settings used | Allow Image Upload, Allow Document/File Upload, and Document Upload Restrictions on Model & Context — the same ones chat uses. |
| Limits | Turn-by-turn always reads images. Natural conversation passes images to the Answering AI model, so they are offered only when that model takes them in a call. Documents are sent as extracted text, cut to about 12,000 characters each (24,000 per turn) in Turn-by-turn and about 8,000 each in Natural conversation with a built-in model; the full text is still recorded. |
Related guide: Talk by Voice