A Lightweight Modular n8n Pipeline for Automated Avatar Narration

YouTube channel link: www.youtube.com/@keywordsin2minutes

Why It Started

This project began as a practical way to turn my ongoing AI learning notes into short, repeatable videos without relying on heavy generative tools.

Long-form automation creates too much drift and review overhead, so I focused on a stable, low-tech format that could be produced consistently. The target length 1:30–2:00 keeps each explainer focused, easy to maintain, and suitable for high-frequency publishing.

Design Choices

The visual style follows the same lightweight philosophy as the workflow: simple, calm, and easy to watch repeatedly. The channel uses a soft blue palette, and the avatar is designed as a soft, androgynous female figure—neutral, friendly, and approachable across genders and age groups.

Although the initial design included 10+ expression states, the pipeline proved most stable with 4-5 core expressions, which kept the visuals consistent and reduced rendering errors. Each video uses a limited set of cues, predictable illustration changes, and a stable pose to keep the format maintainable.

To avoid visual fatigue in a largely automated format, the background moves subtly over time and the character “bounces” lightly between expression states. These small motions provide moments of visual refresh—enough to keep the screen engaging without adding distraction or complexity.

How It Works

The pipeline converts a plain text script into a synchronized video using lightweight components orchestrated through n8n. GPT-4o-mini TTS generates narration, Whisper produces word-level timestamps, and a custom Node.js sync engine reconciles the two—correcting pacing differences, handling skipped words, and producing a clean timeline for FFmpeg.

Real constraints shaped this architecture. TTS pacing varies between runs, Whisper occasionally misaligns tokens, and timestamp rounding introduced subtle frame drift when combined with FFmpeg. Rendering too many micro-clips also led to concatenation failures. To stabilize the workflow, the system retries unreliable steps, normalizes timing data, merges identical visual states, and logs intermediate results into CSV for quick inspection.

Despite these limitations, the pipeline remains fast and inexpensive to iterate on. A full script-to-video pass takes about 3 minutes on an old intel macbook, making debugging cheap, reruns effortless, and TTS/Whisper token usage minimal.

What’s Next

The current system produces clean two-minute explainers with synchronized narration, subtle motion, and controlled expressions. It remains fully creator-ready—any part of the automated workflow can be replaced with human input, including voiceovers, illustrations, or timing adjustments.

Future improvements will explore more dynamic elements such as lightweight motion graphics, animated filters, and playful meme-style overlays to enhance engagement. Expanding the expression set is also a planned direction, allowing the avatar to react more naturally without compromising stability.

All future iterations will follow the same principle:

Add personality without sacrificing the lightweight, reliable foundation that makes continuous creation possible.