Generate a talking-avatar video podcast episode from one topic idea.
Mix voices from different TTS providers into a single leveled conversation.
Produce lip-synced AI avatar clips for short-form social content.
Prototype an explainer video with two hosts without filming anyone.
| jellypod-inc/podframes | codee-sh/payload-training-app | itsfriedpotato/potatopad | |
|---|---|---|---|
| Stars | 10 | 10 | 10 |
| Language | TypeScript | TypeScript | TypeScript |
| Setup difficulty | moderate | moderate | hard |
| Complexity | 3/5 | 4/5 | 5/5 |
| Audience | vibe coder | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Needs Node 22+, pnpm, ffmpeg, and paid API keys for Speechbase, Gemini, and Replicate.
Podframes is a TypeScript tool that turns a single topic into a two-host video podcast, complete with talking AI avatars whose mouths move in time with the words being spoken. You give it a topic, and it writes a script for two hosts using Gemini, then sends that script to a service called Speechbase, which can mix voices from different providers, such as Google, ElevenLabs, OpenAI, or Cartesia, into one leveled conversation with word-by-word timing information. Those word timestamps are what make the rest of the pipeline work. An avatar image is generated for each host with Nano Banana 2, and then each line of dialogue is animated against its real audio using a lip-sync video model, either P-Video Avatar by default or a higher-fidelity alternative called LTX-2.3. Short video clips for background visuals, called b-roll, are pulled in based on the transcript, and everything is composed and rendered into a finished MP4 with captions using a tool called HyperFrames. You can run podframes from the command line or through a web-based studio that walks you through the same steps visually: picking voices, watching the render happen live, and checking the cost estimate before spending money on the more expensive image and video steps. The studio lets you edit a single line of a script and only regenerate that one line's audio and clip, rather than redoing the whole episode, and it warns you before any edit would throw away paid work. There are eight different caption styles and three visual treatments, ranging from a simple talking heads cut to a more graphic heavy, fast paced version, and switching between them does not require paying for new audio or video. To run it yourself you need Node 22 or newer, pnpm, and ffmpeg installed, plus paid API keys for Speechbase, Gemini, and Replicate. The project is licensed under Apache 2.0. Costs are described as rough estimates in the README, generally a few cents per short episode for the speech step, with the image and video steps costing more depending on length and provider choice.
Turns a single topic into a two-host AI video podcast, with talking avatars lip-synced to mixed, word-timed speech.
Mainly TypeScript. The stack also includes TypeScript, Node.js, Gemini.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly vibe coder.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.