Skip to main content
Convo Mode is Pria’s live two-way voice conversation. Your words are transcribed as you say them, the AI model works out a reply, and the reply is spoken back — with an optional video avatar on top. This page covers the admin setup. Most settings are in Instance Settings › Voice; a few live only in the admin editor (Admin › Digital Twins, edit the twin). For the member experience see Convo Mode.

​
What Convo Mode Is

Once enabled, members with access see a Convo button in the toolbar. A session:
  • Captures microphone audio and transcribes it.
  • Shows a running transcript (optional).
  • Runs the Digital Twin’s instructions, files and tools — or the assistant’s, when one is running — on every turn.
  • Speaks the reply in the provider’s voice, optionally with a video avatar.
Your members’ audio does not pass through Pria’s servers during the call. Pria hands the browser a short-lived pass at the start, then the browser talks directly to the voice and speech-recognition providers. Your permanent API keys never reach the browser.

​
Who Can Use Convo Mode

These switches are in Instance Settings › Voice, under Convo (Speech-to-Speech). All but the first appear once Convo is on.
Master switch for the Digital Twin. When off, no one sees the Convo button. Default: off.
Only admins can start Convo Mode. Use it while you test a provider, tune voices or set up an avatar. Default: on — turn it off when you are ready for everyone.
Off by default: every assistant of the Digital Twin can be used in Convo Mode, and a conversation that runs an assistant keeps it when the member starts talking. Turn it on to limit Convo Mode to the assistants you check in the list that appears; members can then switch between them during a call. With the switch on and nothing checked, every assistant stays allowed.
Members can type during a call as well as speak — useful in quiet rooms or for accessibility.
Shows the live transcript in full-screen Convo Mode.
Opens Convo directly in a full-screen avatar view instead of the smaller panel.
Once a session reaches the full-screen avatar view, people stay there for the whole conversation — the collapse button and the Escape key are turned off, and ending the call is the only way out. Use this for kiosk screens and embedded experiences. Default: off.
New visitors land in the immersive voice conversation (with a tap-to-start step) every time they sign in, instead of the chat. Default: off.
Slides a panel of your own content (text, images, links) in beside the full-screen avatar. Choose Drawer Side and build the panel with Open Drawer Builder.
For a Digital Twin embedded on another website through the Pria SDK. The first lets that page receive the live transcript; the second lets it send instructions that steer what the twin says next (those turns are billed like any other). Both are off by default — turn them on only for a website you trust.
Speech Recognition Language (Deepgram Flux) sets what the twin can hear: English only (default) or Multilingual (German, English, Spanish, French, Hindi, Italian, Japanese, Dutch, Portuguese and Russian). It applies to avatar conversations, embeds and meeting bots. A public link that opens straight into Convo without signing in is the Convo Web Access Link, under Instance Settings › Invite › Guest Access — see Guest Twin Links.

​
Choosing a Provider

Pick the provider with Speech-to-Speech Model in Instance Settings › Voice. The model you choose decides the provider and which voice options appear below it.

OpenAI Realtime

Low latency, built-in OpenAI voices. gpt-realtime-2.1 is the platform default; gpt-realtime-2.1-mini costs about a third as much.

OpenAI GPT-Live

Listens and speaks at the same time, so conversation flows naturally and interruptions feel smooth. Your Digital Twin’s own model, files and tools supply every answer.

Gemini Live

Google’s live voice, with 30 voices. gemini-3.8-live is the newest, at the same price as Gemini 3.1 Flash Live.

xAI Realtime

Grok voice (grok-3-fast) with five voices: Eve, Ara, Rex, Sal and Leo.

Anam Avatar

A lifelike video avatar. Anam renders the face and voice; Pria writes every answer, so your instructions, files and tools still apply.

Cartesia Voice

Fast, expressive voices from a large library (including your own cloned voices). Pria transcribes, answers and runs the tools; Cartesia speaks.

ElevenLabs

Voices from an ElevenLabs agent you configure in the ElevenLabs dashboard.

​
Per-Provider Configuration

On gpt-realtime-2 and newer, Voice Reasoning Effort also applies. The twin’s OpenAI API Key is used when set; otherwise Praxis provides access.
Choose gpt-live-1. GPT-Live greets, acknowledges and keeps the conversation going on its own, and hands every question that needs facts, files, memory or an action to your Digital Twin, then speaks the answer. While a longer lookup runs, it keeps the person company instead of going silent.Pictures can’t be shown to GPT-Live during a call. If many GPT-Live conversations run at the same moment, a new one may be refused with “Convo Mode is busy right now — try again in a minute.”
The twin’s Gemini API Key (admin editor, API Keys) is used when set. Google closes a Gemini voice connection after a while; Pria reconnects and carries the conversation on, so a long call keeps going.
The twin’s xAI API Key (admin editor, API Keys) is used when set.
Praxis sets up the Anam account behind the avatar. In Instance Settings › Voice:Apply the avatar standard. Right under Anam Response Model, this button sets the twin to the configuration we recommend for avatar twins. It says how many settings differ, and lists every change for you to confirm before anything is saved; when the twin already matches, it is greyed out. The standard also opens Convo to every member (Convo admin only off), starts in the full-screen immersive view, and turns Enable Moderation and Enable Subagents off — read the list before you confirm. Everything can still be changed by hand afterwards.In the Studio UI’s admin editor, the Video avatar card also offers Expressive voice and face (emotion-aware delivery, off by default) and, under Show advanced, Fill the box in immersive view (the video fills its box edge to edge, cropping instead of showing a blurred border; off by default), Portrait avatar on the web, Portrait avatar on phones and Avatar sound (EQ & compressor).
Set in the admin editor’s Cartesia Voice section:Voice Response Model (Instance Settings › Voice) picks the model that answers spoken turns; empty = the Digital Twin’s conversation model. Praxis sets up the Cartesia account.
Set in the admin editor, Configuration and Integrations tab:ElevenLabs also needs the Chat Completion endpoint turned on, so the agent can call your Digital Twin for its answers — see Chat Completions Integration.

​
Voice Reasoning Effort

A spoken turn can use a different amount of thinking from a typed one, and usually should. Voice Reasoning Effort (in the admin editor next to Speech-to-Speech Model) sets how long the model works before it starts speaking, on the same scale as Reasoning Effort — None, Low, Medium, High, Max. The default is None. It applies to Anam, ElevenLabs, Cartesia and GPT-Live, and to OpenAI voice on GPT-Realtime-2 and newer. Older OpenAI voice models ignore it. An assistant can set its own with STS / Voice Reasoning Effort. Keep it low. In a spoken conversation thinking time is silence, and silence reads as the twin having failed to hear you.
Low is the measured starting point for avatar twins (it is part of the avatar standard). Raise it only for a twin whose spoken questions genuinely need working through, and listen to the pause before you keep it.

​
Tools During Voice Sessions

A voice conversation offers the AI the tools made for voice, plus any tools an assistant was limited to. The twin waits at most a few seconds for a tool: a longer job, such as saving a file, keeps running in the background while the conversation continues, and the twin gives the result on the next turn (Gemini Live waits as before). When a connector tool needs approval, most providers show the approval card during the call; on ElevenLabs the twin asks the member to try again from the chat.

​
Testing Your Setup

1

Keep Convo Mode admin-only while testing

Leave Convo admin only on while you tune voices, listening and the avatar. Only admins see the Convo button.
2

Open Convo Mode

Sign in as an admin, open a conversation, and start Convo Mode.
3

Verify the basics

Check that the avatar (if any) appears, your voice is transcribed, the answer comes in the chosen voice, and the twin’s tools still work.
4

Tune listening if needed

On OpenAI Realtime, if the twin interrupts people mid-thought, set Voice Activity Detection Eagerness to Low; if replies feel slow, try High. On Gemini Live, use Interruption Sensitivity.
5

Open it to everyone

When you are satisfied, turn Convo admin only off.

​
Cost Considerations

Convo Mode is billed per minute by every provider. Costs vary widely:
  • OpenAI Realtime and Gemini Live are billed on input/output audio tokens per minute by the provider.
  • ElevenLabs charges per character of speech plus session minutes.
  • xAI Realtime is billed on audio minutes.
  • Cartesia charges for speech output on top of the answering model.
  • Anam charges per minute of active avatar time on top of the answering model.
  • OpenAI GPT-Live charges per second of conversation (about $0.05 a minute) on top of the model that answers.
If you bring your own provider API keys, the provider bills your account for those minutes; Pria still meters the session in credits (BYOK customers buy credits at a discount through our sales team).

​
What your members are charged

What comes off a Digital Twin’s credit balance is simpler: a one-to-one Convo Mode session — avatar or voice-only — is metered on the tokens of the conversation, totalled across the session and charged in credits when the session ends. There is no separate per-minute avatar charge in Convo Mode, so an avatar session and a voice-only session of the same length cost about the same in credits. The one exception is OpenAI GPT-Live: its conversation time is added to the credits as well, at OpenAI’s per-minute price plus the platform margin, on top of the tokens of the model that answers. A ten-minute GPT-Live conversation adds roughly 7 credits for the voice time. In practice people stay in conversation with a face on screen far longer than in a chat box, so avatar twins use up a pool faster. Watch the pool, not the per-minute rate. See Credits for the metering rules and Credit Management for per-user caps.
This page covers one-to-one Convo Mode. When a Digital Twin is sent into a meeting, billing works differently — see Meeting Agent.
Anam costs the most per minute of provider time. If you don’t need a visible persona, OpenAI Realtime or Gemini Live with a good voice costs a fraction as much.