Skip to main content

A talking, interruptible AI presenter is now open source

By Alan Magliocca · November 14, 2026

Quick answer: On September 10, 2026, HeyGen released an open-source starter, under the MIT license, that connects OpenAI's interruptible GPT-Live-1 voice model to a HeyGen LiveAvatar with visuals on screen. Treat it as a working blueprint for a one-to-one AI presenter, not a webinar platform, and list what you would still have to build before putting one in front of buyers.

KEY TAKEAWAYS
  • HeyGen's liveavatar-gpt-live-demos repo connects GPT-Live-1, a full-duplex voice model, to a HeyGen LiveAvatar, and is MIT-licensed.
  • It needs a LiveAvatar API key and an OpenAI key with GPT-Live access. GPT-Live-1 costs $0.05 a minute, billed per second.
  • HeyGen calls it "a starter, not a deployment" and says to add authentication before exposing it publicly.
  • The demo is one person talking to one avatar. A webinar also needs registration, reminders, a shared audience, chat, replays and sales tracking.
  • Anyone running an AI presenter should tell attendees it's AI at the start.

What happened

On September 10, 2026, HeyGen published liveavatar-gpt-live-demos, an open-source reference integration that connects OpenAI's GPT-Live-1 speech model to a HeyGen LiveAvatar, an animated face that talks in real time. HeyGen announced it in its September 2026 release notes, the same day OpenAI put GPT-Live-1 in its API.

What the code does, according to HeyGen's release notes and the repo's README:

  • Voice: GPT-Live-1 is a full-duplex model, which means it listens and speaks at the same time. OpenAI's model page describes "smooth interruption handling" and lets it hand reasoning and tool use to a backend model.
  • Face: a HeyGen LiveAvatar, with LiveKit carrying the avatar's video and audio to the browser.
  • Visuals: when the model calls a tool, an animated overlay appears on screen beside the avatar. The Japanese tutor demo shows a term card as it says each word and recaps the words taught so far.
  • Persona: defined in two markdown files. HeyGen says adding a new on-screen visual takes three small edits.
  • Listening: there's no push-to-talk button. The microphone streams all the time and the model decides when you've finished speaking.

HeyGen's release notes list a language tutor, a poker coach and a customer support agent among the demos. A second repo lets the avatar request motion graphics mid-conversation, which a background Claude Code session writes in about 30 seconds to two minutes while the avatar keeps talking.

Both repos are MIT-licensed. The bundled GSAP animation library carries its own license. To run it you need Node 20.12 or later, a LiveAvatar API key and an OpenAI API key with GPT-Live access. HeyGen is direct about its status: the README calls it "a starter, not a deployment" and says to add authentication and review a hardening list before public use.

What it means for people selling with webinars

An AI presenter you can interrupt mid-sentence is now a reference architecture rather than a research project. The full pattern, voice plus face plus on-screen visuals, is public code that a developer can install and run with two API keys. That changes the conversation with vendors. When someone sells you an "interactive AI host," you can ask how theirs differs from a free starter.

It doesn't mean you can run a sales webinar on it. The demo is one person talking to one avatar in one browser. A sales webinar is a different shape: hundreds of registrants, reminders, a shared room, a chat, an offer, a replay and a record of who bought. None of that is in the repo, and HeyGen doesn't claim it is.

The likely near-term use for webinar sellers is narrower. Think of a one-to-one AI conversation before or after the event, such as a pre-event Q&A on the registration page or a post-event follow-up, rather than an AI that presents to the whole room.

What to do this week

  1. Decide what you want an AI presenter to do. Present to everyone, answer questions one-to-one, or both. Write it in one sentence.
  2. If you have a developer, have them run the starter for one hour with your own test keys. Interrupt it on purpose and note how quickly it stops and responds.
  3. Price one session. GPT-Live-1 voice is $0.05 a minute, so a 10-minute one-to-one conversation is $0.50 in voice before the avatar, backend model and hosting. The full math is in what an interruptible AI host costs.
  4. Write your disclosure line before anything goes live. Use a plain-English disclosure line for AI-hosted webinars as a starting point.
  5. Ask your current vendors whether their AI presenter can be interrupted, and how many attendees can talk to it at once.

How BrightStage AI helps

BrightStage AI doesn't use this open-source stack and doesn't offer a voice host attendees can interrupt. What it has now: a video creator that builds a script and slides with AI presenters, and a chat assistant that answers attendees in the event chat during live and automated events. Around that sit the parts the starter leaves out, such as A/B/C split-tested registration pages, reminders by email, text and web push (confirmed on desktop and Android), and Encore replays.

Build or buy: what the starter leaves out

If you're weighing a custom AI presenter built on this code against a platform, here's what still has to exist before buyers see it:

  • Authentication and abuse limits. HeyGen's first hardening step. Without it, anyone with the link can run up your bill.
  • Many viewers at once. The starter connects one browser. Serving an audience needs streaming, and every private conversation is a separate session. OpenAI's model page caps concurrent sessions at 50, 300 or 500 depending on account tier.
  • Registration and reminders. Who's attending, when, and how they get the link.
  • Chat and moderation. What happens when someone types instead of speaks, or says something abusive.
  • Recording and replays. The starter doesn't save the session for later viewers.
  • The offer and checkout. A link, a deadline and a record of purchases.
  • Disclosure and consent. A clear AI notice, and written consent if the avatar or voice is based on a real person.
  • Upkeep. Two vendor APIs that will change. Someone has to keep the code current.

How fast does an interruptible presenter need to be?

Speed is what gives an AI host away. OpenAI reported turn-taking latency of about 0.8 seconds for GPT-Live-1, down from 1.41 seconds for the previous model. That is OpenAI's own figure, and an avatar adds its own delay on top. Test the full chain, voice plus face, on an ordinary home connection, not just the voice model on its own. We compare response times in 0.6 vs 0.8 seconds. For a plain explanation of the parts behind any AI-hosted event, AIVX Labs covers what an AI webinar host is and how it works.

Sources

AIVX LABS ACCOUNT VERIFICATION
Get AIVX Labs verified for $1: RICO, your AI operator, and 35+ AI tools, with 7 days of full access.
Get AIVX Labs Verified for $1 →

Frequently asked questions

Is HeyGen's AI presenter code free to use?

The code is MIT-licensed, so you can use and change it. Running it isn't free: GPT-Live-1 voice costs $0.05 a minute, and LiveAvatar is a separate HeyGen service with its own account and API key.

Can I run a webinar with this open-source starter?

Not on its own. It connects one person to one avatar in a browser. Registration, reminders, a shared audience, chat, replays and sales tracking would all have to be built or added.

Can attendees interrupt the avatar?

GPT-Live-1 is a full-duplex model that OpenAI describes as handling interruptions, and the starter streams the microphone continuously with no push-to-talk. Test it yourself to see how quickly it stops and responds.

Related reading

Written by Alan Magliocca, founder of BrightStage AI and AIVX Labs (Podavinci LLC). Written with AI assistance and checked against the linked sources. Follow along: LinkedIn · YouTube · TikTok · X · Facebook · Instagram · Skool community