One prompt. A full animated podcast show — dialogue, voices, lip sync, cameras, final cut. Fully automated, end to end.
English · 简体中文 · Live Demo · Documentation
Just type one sentence — "Make a podcast about coffee culture" — and Blabber writes the dialogue, casts two hosts, voices every line, syncs every lip movement to the audio waveform, directs the cameras, and renders the final cut. You describe the show. The agents make it.
- [2026-07-21] 🎉 Blabber officially launched on GitHub!
- [2026-08-07] 🚀 Released the first version with the Zoo scene and two video-ready hosts: Gaga and Awang.
- [2026-08-20] ✨ Major update since August 7: added Luna, Milo, and more scenes; expanded dual-host voice controls and script, subtitle, and timeline editing; introduced parallel rendering, clearer generation feedback, LAN deployment, and improved audio/video sync stability.
Six 12-second excerpts from animated podcast episodes generated with Blabber.
🧚 Fairy Tale
fairy-tale.mp4 |
💞 Modern Relationships
modern-relationships.mp4 |
🗣️ Low-EQ Behaviors
low-eq-behaviors.mp4 |
🏛️ Government Policy Explainer
government-policy.mp4 |
☕ Coffee Culture
coffee-culture.mp4 |
🚗 New Car
new-car.mp4 |
Blabber is an AI production platform built for podcast video creation. You don't record, animate, or edit anything — you just describe the show you want, and a team of AI agents orchestrates the entire journey: topic planning → dialogue script → voice casting → speech synthesis → lip sync → camera direction → editing → final render.
It's not just a one-shot video generator. Blabber turns every show into an editable project: after generation, you can keep refining anything — a line of dialogue, a host's voice, a camera cut, a scene — through the built-in AI Copilot chat or directly on the clip-based timeline.
| 1. Describe the show Enter a topic or upload source material.  |
| 2. Review the script Check the generated outline, dialogue count, and estimated length.  |
| 3. Choose the hosts Pair the show with characters and matching voices.  |
| 4. Select a scene Pick a visual style and preview the composition.  |
| 5. Preview the video Review the finished episode and its render details.  |
| 6. Export the result Export the video, cover image, or audio track.  |
- End-to-end generation — turn one prompt or document into a two-host script, voices, animation, captions, camera cuts, and a finished episode.
- Natural two-host conversation — generate character-matched voices, reactions, follow-up questions, and editable line-by-line audio clips.
- Automated visual production — frame-accurate lip sync, speaker-aware camera direction, and final FFmpeg composition.
- Conversational editing — revise scripts, voices, shots, and timing through AI Copilot or the clip-based timeline, with candidate versions and rollback.
- Flexible presentation — mix bundled characters and scenes, then export captioned videos for widescreen, vertical, or square platforms.
Prerequisites:
- Git
- Node.js
>=22.13.0 - Python
>=3.9 ffmpegandffprobeavailable onPATH, withlibvpx-vp9support- About 1 GB of free disk space for runtime caches and output. The bundled VP9 Alpha action assets add about 52 MB to the regular Git clone; Git LFS is not required.
First-time setup:
git clone https://github.com/GML-MMGroup/Blabber.git
cd Blabber
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r mvp/requirements.txt
cp mvp/.env.example mvp/.env
cd site
npm ciOn Windows PowerShell, activate with .venv\Scripts\Activate.ps1 and create
the configuration file with Copy-Item mvp/.env.example mvp/.env.
If PowerShell blocks npm.ps1 because of its execution policy, use npm.cmd for the documented npm commands (for example, npm.cmd ci and npm.cmd run dev). Verify the prerequisites before setup with node --version, python --version, ffmpeg -version, and ffprobe -version. Video segments are rendered concurrently: action video defaults to at most 4 workers and lip-sync video to at most 2. Set BLABBER_VIDEO_WORKERS in mvp/.env to a positive integer to override both; lower it if CPU, memory, or GPU memory is limited.
The web UI uses Volcengine PodcastTTS to return two-host dialogue slices, per-slice MP3 files, and the complete MP3 in one API request, ready for MP4 generation. After setup, start the two local services in separate terminals.
Terminal 1 — backend:
cd Blabber
source .venv/bin/activate
cd site
npm run dev:mvpTerminal 2 — frontend:
cd Blabber
cd site
npm run devOpen the local URL printed in the terminal (normally http://localhost:3000). Configure the Doubao Speech PodcastTTS App ID and Access Token under service settings. To use Seedream or Seedance, additionally run python -m pip install -r mvp/requirements-ark.txt and set the optional ARK_API_KEY in mvp/.env. This credential file is ignored by Git and must not be committed.
Enter a topic or select a document, then click Generate script and audio. Once audio is ready, continue with MP4 generation using the bundled Gaga and Awang characters in the Zoo scene.
The site/ directory is configured for OpenAI Sites. From that directory, run npm ci and npm test before publishing through Sites. The checked-in .openai/hosting.json identifies the existing Sites project; do not replace its project_id. Optimize large tracked static assets before publishing; oversized images or an excessive source payload may cause the Sites source push to be rejected.
Sites publishes the web UI and its Worker only. The Python service in mvp/ performs PodcastTTS calls, media persistence, FFmpeg processing, and MP4 rendering, and is not included in the Sites deployment. The development proxy in site/vite.config.ts forwards /api/mvp and /mvp-media to 127.0.0.1:8787 only while running locally. A functional public deployment therefore requires a separately hosted Python backend plus a production reverse proxy (or equivalent same-origin routing) for those two paths. Never expose mvp/.env or put API credentials in frontend environment variables.
POST /api/mvp/document-jobs uses PodcastTTS document mode (action=0) and accepts a webpage or downloadable document URL, long text, or a local file. The UI accepts .txt, .md, .html, .json, .csv, .docx, and .pdf files up to 20 MB and shows the selected file name and size:
curl -X POST http://127.0.0.1:8787/api/mvp/document-jobs \
-H 'Content-Type: application/json' \
-d '{"input_url":"https://example.com/article","topic":"Article explainer"}'Alternatively, use {"input_text":"Document body...","topic":"Document explainer"} or upload a Base64 file with {"file_name":"report.pdf","file_base64":"...","topic":"Report explainer"}. Provide exactly one of input_url, input_text, or the file upload fields; topic is optional. Scanned PDFs must be OCRed first. The endpoint returns 202 with a job ID. Poll GET /api/mvp/jobs/{id}; the completed job includes episode.turns, clips[].audio_url, audio_url, and provider_audio_url.
Jobs are persisted in mvp/output/jobs-history.json; read recent records from GET /api/mvp/history. The UI can restore saved scripts, audio slices, and videos, while identical new inputs reuse completed results without another paid request. During generation, GET /api/mvp/jobs/{id}/events streams accumulated script and audio slices over SSE.
⭐ If you find Blabber useful, please give us a star!
Made with ❤️ by the Blabber Team