Open-Source digital-human-koubo Turns a Script Into a Polished AI Presenter Video

The open-source digital-human-koubo project turns a presenter image and script into a vertical or landscape video with lip-synced speech, normalized audio, captions, timed graphics, and creator-style editing.

Berryxia, also known as andyhuo520, released the MIT-licensed project as an Agent Skill. Its SKILL.md file tells a compatible coding agent how to coordinate the included Python, shell, and Remotion scripts. Remote generation runs through a Model Context Protocol server, while local tools validate, edit, and render the result.

A video pipeline packaged as code

The skill models each video as a reproducible build with nine stages:

1. Analyze the topic and recommend two production styles.
2. Create a structured script and divide it into segments of up to 15 seconds.
3. Generate presenter clips through Grok Video 1.5, Wan 3.0, or OmniHuman.
4. Transcribe each clip with Whisper and compare the transcript with the script.
5. Sample four frames from each segment to detect unwanted burned-in subtitles.
6. Match voice tone, sibilance, room sound, and loudness across clips.
7. Apply captions, graphics, sound effects, and camera treatments.
8. Validate component names, timings, and source files.
9. Render the assembled video with Remotion.

A persona.json file stores the presenter’s face, clothing, voice description, and prompt template. Reusing that configuration gives each generated segment the same on-screen identity, even when a long script requires dozens of separate model calls.

The audio pass applies one-third-octave spectral matching, de-essing, shared small-room reverb, and loudness normalization. In the author’s tests, variation in sibilance levels fell from 13.2 dB to 5.9 dB, reducing audible jumps between independently generated clips.

Five formats share one editing grammar

Five built-in styles define the aspect ratio, duration, pacing, and visual treatment for common video formats: