video-metadata

Automated pipeline for preparing YouTube metadata from finished videos and blog articles. Generates titles, descriptions, tags, timestamps, and SRT subtitles in both English and Russian.

🏭 The Content Factory

video-metadata is stage 3 of an automated pipeline that turns a blog article into a published YouTube video — no API keys, driven end-to-end through logged-in browser sessions (Camoufox) and local media tooling.

#	Stage	Repo	What it does
1	Generate	gaia	Drive NotebookLM / Gemini / Flow from a logged-in session → audio overview + slide deck
2	Build	video-maker	Audio narration + slide-deck PDF → synced MP4 (+ SRT, thumbnail)
3	Describe	video-metadata ⬅ this repo	Video + article → YouTube title / description / tags / chapter timestamps
4	Publish	video-publisher	Drive YouTube Studio → upload with metadata, channel switch, visibility

Flow: article → gaia → video-maker → video-metadata → video-publisher → YouTube (the published video is then embedded back into the blog article).

What It Does

Given a finished video file and a corresponding blog article slug, the pipeline:

1. Extracts audio from the video using FFmpeg
2. Transcribes speech to text using OpenAI Whisper
3. Generates SRT subtitles for YouTube upload
4. Builds chapter timestamps by matching article section headings to transcription segments
5. Researches YouTube tags from article content (headings, key phrases, frontmatter tags)
6. Generates metadata (title, description, tags, timestamps) in both EN and RU — as JSON and human-readable TXT

Requirements

• Python 3.9+
• FFmpeg installed and available in PATH
• OpenAI Whisper (pip install openai-whisper)
• yt-dlp (optional, for tag research via YouTube search)

Python Dependencies

pip install openai-whisper

Project Structure

video-metadata/
├── scripts/
│   ├── prepare_metadata.py   # Main pipeline script
│   └── run.sh                # Batch runner for multiple videos
├── input/                    # Place source video files here
├── output/                   # Generated metadata, SRT files
└── temp/                     # Intermediate files (audio, subtitles JSON)

Usage

Single Video

python scripts/prepare_metadata.py \
    --video input/my_video.mp4 \
    --slug my-blog-article-slug \
    --output-dir output/my-blog-article-slug \
    --whisper-model base

Batch Processing

Edit the VIDEO_MAP in scripts/run.sh to define video-to-slug mappings, then:

bash scripts/run.sh

CLI Options

Option	Description	Default
--video	Path to input video file	required
--slug	Blog article slug (used to find .en.md / .ru.md articles)	required
--output-dir	Directory for output files	required
--temp-dir	Directory for intermediate files	temp/<slug>
--whisper-model	Whisper model size (tiny, base, small, medium, large)	base
--skip-transcribe	Reuse existing subtitles.json from temp dir	false

Output Files

For each video, the pipeline generates:

File	Description
<slug>.srt	YouTube-ready SRT subtitles
<slug>_metadata.json	Structured EN metadata (title, description, tags, timestamps)
<slug>_metadata.txt	Human-readable EN metadata for copy-paste into YouTube Studio
<slug>_ru_metadata.json	Structured RU metadata
<slug>_ru_metadata.txt	Human-readable RU metadata

How It Works

Timestamp Generation

Instead of fixed-interval timestamps, the tool matches blog article section headings against Whisper transcription segments using word overlap + bigram scoring. This produces accurate chapter markers that correspond to actual topic transitions in the video.

Tag Generation

Tags are extracted directly from article content:

• Section headings split at separators (parentheses, colons, dashes)
• Key phrases from the article description (2-3 word meaningful combinations)
• Frontmatter tags and their pairwise combinations
• Brand tags (marketmaker, marketmaker_cc) are always appended

Blog Article Integration

The tool expects blog articles in Markdown with YAML frontmatter at a configurable path (default: ../../marketmaker-cc-landing/src/content/blog/). Articles should have title, description, and tags fields, plus section headings (##) in the body.

Related Projects

• gaia — generates the audio + slides upstream
• video-maker — generates the video itself from audio + PDF slides
• video-publisher — uploads the video (with this metadata) to YouTube

License

MIT