Skip to main content
Pria’s Convo Mode lets users talk to a Digital Twin in real time — speech in, speech (and optionally an animated avatar) back out. Under the hood, several different vendors can power that experience. This page compares them so you can pick the right one for each Digital Twin.
You don’t have to pick just one. Different Digital Twins can use different voice providers, and on the providers that offer a voice list, each user can pick their own voice — unless you fix one for everyone.

​
What Realtime Voice Is

Realtime voice is a live, two-way audio conversation between a user and the Digital Twin. Unlike recording a clip and waiting for a transcript, realtime voice streams audio in both directions:
In words: you speak → your words are transcribed as you say them (speech-to-text, STT) → the AI works out a reply → the reply is spoken back as it is generated (text-to-speech, TTS) → you hear it. A few vendors (Gemini Live, OpenAI Realtime, xAI Realtime) handle listening, thinking and speaking inside one voice model, which responds fast and sounds natural. Others (ElevenLabs, Cartesia) chain separate listening and speaking components around your Digital Twin’s own model for premium voice quality. OpenAI GPT-Live listens and speaks at the same time, so interruptions feel smooth, and hands every real question to your Digital Twin’s own model, files and tools. The avatar provider (Anam) adds an animated, lip-synced video on top of the voice. In every case the audio itself travels straight between the listener’s browser and the voice vendor. Pria sets up each call, never handing your vendor key to the browser, and then stays out of the audio path, so sound never queues behind Pria’s servers.

​
Provider Matrix

​
Choosing a Provider

Recommendation: Gemini Live or OpenAI Realtime.Native audio models give the lowest latency, which matters most for back-and-forth dialogue with learners. Gemini Live shines for multilingual classrooms; OpenAI Realtime has the broadest model coverage and stock voices.
Recommendation: ElevenLabs.Premium voice quality is what customers remember. ElevenLabs also supports a published voice agent that can be embedded in support widgets via the Chat Completions API bridge.
Recommendation: Anam Avatar.A lifelike animated avatar dramatically raises perceived presence — useful for product demos, role-play training, and onboarding videos. Pair it with your conversation model of choice for the brains, and Anam handles voice + lipsync + video.
Recommendation: Anam Avatar or OpenAI Realtime with transcripts enabled.For users who benefit from a visible speaking face (lip-reading, comprehension support), Anam adds an avatar. For audio-only accessibility, enable the live transcript so users can read along.
Recommendation: OpenAI Realtime or Gemini Live.Native-audio models handle code-heavy content and technical terms well. An avatar can struggle to render code snippets in spoken form.

​
Tradeoffs

Voice quality is highest with ElevenLabs and Cartesia (dedicated TTS engines) and with the Anam avatar when paired with a quality voice. Native-audio models (Gemini Live, OpenAI Realtime, xAI Realtime) sound natural and conversational but don’t expose voice cloning. Latency is lowest with OpenAI Realtime via WebRTC, followed closely by Gemini Live and xAI Realtime over WebSocket. ElevenLabs adds a small additional hop because Pria acts as the brain via Custom LLM. The avatar provider adds video rendering on top. Cost depends on whether you bring your own vendor key (BYOK) or use Pria-included credits. In Convo Mode every provider is charged for the model that answers; GPT-Live adds a per-second charge for the conversation time. An Anam avatar adds no per-minute charge in Convo Mode, but sent into a meeting it is billed per minute — see Anam Avatar. Tools during a call. On Anam, Cartesia, GPT-Live, OpenAI Realtime and xAI, the twin waits at most about 6 seconds for a tool; a longer job keeps running in the background and the twin tells you the result on your next turn. Gemini Live waits for the tool to finish. How each provider asks before a connector tool runs is in MCP in Real-Time Voice Conversations.

​
Where Avatars Matter

Avatars are not always the right answer — they add cost and a layer of visual polish that can distract from purely informational use cases. Consider an avatar when:
  • Sales demos — a face on screen makes the product feel “alive” and quotable in marketing.
  • Training & onboarding — embodied presence raises engagement and retention for long-form content.
  • Accessibility — visible mouth movement supports lip-reading and comprehension.
  • Marketing landing pages — a greeter avatar is more memorable than a chat widget.
Skip the avatar for fast Q&A, code help, or text-heavy conversations where users want to read along.
You can also put an Anam-powered avatar on an external website — visitors talk to it without signing in to Pria. See the Avatar Embed Widget.

​
Setup Overview

Each provider has its own setup page with detailed steps:

ElevenLabs Voice Agent

Connect an ElevenLabs ConvAI agent to Pria as a Custom LLM, deploy embeddable widgets.

Gemini Live Voice

Google’s native-audio WebSocket API for low-latency multilingual voice.

Anam Avatar

Lifelike animated avatar driven by your choice of conversation model.

OpenAI Realtime, GPT-Live and xAI

Configure built-in realtime providers from the Admin guide.

Cartesia Voice

Expressive voices (including clones) with Pria doing the listening and thinking — configured from the Admin guide.
At a high level, every setup follows the same order:
  1. Turn Convo on — on Configuration and Integrations, under Realtime / Convo, tick Enable Convo (Speech-to-Speech).
  2. Choose the provider — on Personalization and AI Models, set Speech-to-Speech Model to one of that provider’s models. The provider-specific settings appear only after this.
  3. Fill in what that provider needs — nothing extra for OpenAI (Realtime or GPT-Live), Gemini or xAI when Praxis AI’s platform keys are in use (add your own key under API Keys if you want the usage on your own account); an agent ID and key for ElevenLabs, plus Enable Chat Completion endpoint; a voice for Cartesia; the avatar for Anam (see Anam Avatar).
  4. Decide on the voice — leave it to each user, or fix one for everyone, where the provider offers a choice.

​
Multi-Provider Strategy

Pria does not force a single voice provider per organisation. Common patterns:
  • One provider per Digital Twin, per-assistant model overrides. The voice provider is a Twin-level choice, but on the providers where Pria does the thinking — ElevenLabs, Cartesia, GPT-Live and Anam — each assistant can override the model used for its voice turns with STS / Voice Conversation Model: a fast model for a snappy Sales assistant, a deeper one for a Research Tutor. (OpenAI Realtime, Gemini and xAI think inside their own voice model, so there’s nothing to override.)
  • Different Digital Twins, different providers. A K-12 tutoring Twin uses Gemini Live for multilingual support; a corporate training Twin uses Anam for an avatar-led experience.
  • Per-user voice choice. Where the provider offers a voice list, users can pick a voice in the Convo widget without changing the provider.
This flexibility lets you optimize for the right tradeoff per use case rather than locking the whole organisation into one vendor’s strengths and weaknesses.

​
Per-User Voice Preferences

On OpenAI Realtime, OpenAI GPT-Live, Gemini Live and xAI Realtime, each user picks a voice in Convo Mode, and Pria remembers their pick. It belongs to the person, so it follows them to every Digital Twin that uses the same provider. The provider itself is an admin decision. Two things turn the selector off:
  • The admin fixed a voice. Setting the Digital Twin’s voice to anything other than User’s Choice gives everyone that voice and hides the selector.
  • The provider carries its own voice. ElevenLabs uses the voice of its agent, Cartesia the voice set on the Digital Twin, and the Anam avatar its own voice — none offers a per-user choice.
This is useful when:
  • Different team members have different voice preferences for accessibility or focus.
  • Users want a male/female/neutral voice without affecting the rest of the Digital Twin.
  • A user is testing voices before recommending a new default to their admin.