🎬 video_transcription

Turn any video into a rich, timestamped transcript — audio + on-screen visuals, merged.

Local. Fast. No API keys required (if you already have Claude Code / Codex / Gemini CLI).

Features · Quick Start · How It Works · Agents · Example Output

---

💡 Why?

Traditional transcription tools give you text only. You lose everything on screen — slides, charts, code, UI screenshots. You end up with a transcript that says "as you can see here..." and no way to know what here is.

video_transcription fixes that by running two passes:

1. Audio is transcribed with mlx-whisper (hallucination-guarded, Metal-accelerated).
2. Frames are pulled via scene detection + hybrid interval rules, then handed to your local AI CLI agent (Claude Code, Codex, or Gemini) which describes each frame with the surrounding audio as context — so it understands what it's looking at.

The result: one interleaved timeline you can read, search, or feed back into another LLM.

✨ Features

• 🎯 Audio + visual in one transcript — not two separate files you have to align yourself
• 🔌 Pluggable AI agents — swap between claude, codex, gemini with one flag
• 🧠 Context-aware frame descriptions — each frame is described knowing what was just said
• ⚡ Metal-accelerated — mlx-whisper on Apple Silicon runs transcription faster than realtime
• 🎛️ Smart frame selection — hybrid mode combines scene-change detection with min/max interval guards, plus bitwise dedup for static scenes
• 🗣️ Speaker diarization (optional) — via pyannote through whispermlx
• 🌐 Multilingual — Whisper auto-detects language (tested on Korean, Indonesian, English; works for 99 languages)
• 📄 Markdown + JSON output — human-readable and machine-parseable
• 🔒 Runs locally — your video never leaves your machine (your CLI agent's cloud calls aside)

🚀 Quick Start

# 1. Clone
git clone git@github.com:treesoop/video_transcription.git
cd video_transcription

# 2. Install common deps
brew install ffmpeg
pip3 install -r requirements.txt mlx-whisper

# 3. Install ONE agent CLI (Claude is the default)
curl -fsSL https://claude.ai/install.sh | bash
claude   # first run: OAuth sign-in

# 4. Verify
python3 video_transcribe.py --check --agent claude

# 5. Transcribe
python3 video_transcribe.py my_video.mp4
# → my_video.md

That's it. Drop in a video, get a transcript back.

⚙️ How It Works

                         ┌─────────────────────────────────────────┐
                         │             video.mp4                    │
                         └────────────────┬────────────────────────┘
                                          │
                ┌─────────────────────────┴─────────────────────────┐
                ▼                                                   ▼
    ┌───────────────────────┐                      ┌────────────────────────────┐
    │   ffmpeg → audio.wav  │                      │  ffmpeg scene detection +  │
    │                       │                      │  hybrid interval guards    │
    │   mlx_whisper         │                      │  + MD5 dedup               │
    │   (hallucination-safe)│                      └──────────────┬─────────────┘
    └──────────┬────────────┘                                     │
               │                                                  ▼
    (opt) whispermlx diarize                          ┌──────────────────────┐
               │                                      │  your local agent    │
               ▼                                      │  CLI (parallel x4)   │
    ┌──────────────────────┐                          │  claude | codex |    │
    │ audio_segments:      │  ◀── audio ±30s context  │  gemini              │
    │  [start, end, text,  │  ─────────────────────▶  │                      │
    │   speaker?]          │                          │  describes each      │
    └──────────┬───────────┘                          │  frame contextually  │
               │                                      └──────────┬───────────┘
               │                                                 │
               └────────────────────┬────────────────────────────┘
                                    ▼
                     ┌──────────────────────────────┐
                     │   timeline merge (by time)   │
                     └─────────────┬────────────────┘
                                   │
                      ┌────────────┴────────────┐
                      ▼                         ▼
                 output.md                 output.json
              (human-friendly)          (machine-parseable)

The clever bit is the audio → context → frame flow. Instead of asking an agent to describe a frame cold, we slice the audio transcript around that moment (±30s by default) and feed it in. The agent stops guessing and starts grounding.

🤖 Pick Your Agent

Agent	How it's called	Cost model	Install
Claude Code (default)	claude -p + Read tool	Claude Pro/Max subscription or API key	curl -fsSL https://claude.ai/install.sh | bash
Codex	codex exec -i <image>	ChatGPT Plus/Pro or OpenAI API key	npm install -g @openai/codex
Gemini	gemini -p "@<image>"	Free tier available + Google API	npm install -g @google/gemini-cli

Swap with a flag:

python3 video_transcribe.py my_video.mp4 --agent codex
python3 video_transcribe.py my_video.mp4 --agent gemini

Every adapter implements the same BaseAgent.describe_frame(image, audio_context, system_prompt) interface, so adding another CLI (Aider, Cursor headless, your own wrapper) is a 30-line file in src/agents/.

📺 Example Output

From an 8-minute product demo video (Claude described each frame):

# demo.mp4 — Transcript

**Duration:** 08:12 · **Frames analyzed:** 9 · **Agent:** claude

---

00:00 Alright, let me walk you through the analytics dashboard we shipped last week.

> 🖼️ **[00:00]** A web dashboard in dark mode. Top navigation reads
> "Overview / Analytics / Settings / Billing". Main panel shows a line
> chart titled "Active Users — Last 7 Days" with the line climbing from
> ~1,200 to ~4,800. Matches the speaker introducing the new dashboard.

00:14 The first thing you'll notice is the analytics panel on the left.
00:22 We pull four key metrics right at the top.
00:30 Conversion, bounce rate, session length, and churn.

> 🖼️ **[00:32]** Same dashboard, now with a left-side analytics panel
> highlighted. Four stat cards visible: "Conversion Rate 3.4%",
> "Bounce Rate 42%", "Avg Session 2m 18s", "Churn 1.2%". The speaker's
> focus matches the highlighted element exactly.

00:48 If you click into any metric you get a drill-down.

> 🖼️ **[00:52]** A modal overlay has opened. Table headers:
> "Source / Visits / Conversions / Rate". Visible rows include
> "Direct 2,341 / 112 / 4.8%" and "Organic 1,892 / 67 / 3.5%".
> Corresponds to the speaker describing clicking a metric to drill down.

The frame descriptions know what the narrator is talking about because the audio context was passed alongside the image.

🎛️ CLI Reference

python3 video_transcribe.py VIDEO_PATH [options]

Flag	Default	Description
--agent	claude	claude | codex | gemini
--mode	hybrid	Frame extraction strategy: hybrid, interval, or scene
--interval	30	Fixed seconds between frames (for interval mode)
--min-interval	10	Hybrid: minimum seconds between kept frames
--max-interval	60	Hybrid: force-insert a frame in gaps longer than this
--max-frames	200	Hard cap on total frames analyzed
--scene-threshold	0.15	ffmpeg scene sensitivity (0–1). Lower = more frames. Live-action: 0.3. Animation/illustration: 0.05–0.1.
--diarize	off	Enable speaker diarization (requires --hf-token)
--hf-token	$HF_TOKEN	HuggingFace token for pyannote
--prompt-file	prompts/default.md	Override the frame-description system prompt
--context-window	30	±seconds of audio context fed to each frame
--format	md	md | json | both
--parallel	4	Concurrent agent invocations
--keep-frames	off	Preserve extracted PNGs (default: deleted after run)
--check	—	Preflight check (verifies all deps) and exit
-o, --output-dir	.	Where to write the transcript

🎨 Custom Prompts

Tailor the frame-description style for your use case. Two recipes ship with the repo:

# Business meeting recordings (Zoom, Meet, etc.)
python3 video_transcribe.py meeting.mp4 --prompt-file examples/meeting_recipe.md

# Lectures / technical tutorials
python3 video_transcribe.py lecture.mp4 --prompt-file examples/lecture_recipe.md

Or write your own — prompts are just plain text. The system prompt is sent to each frame call, so describe what you want the agent to focus on (slides, code, diagrams, product UI, etc.).

🗣️ Speaker Diarization

Need "who said what"? Add --diarize:

python3 video_transcribe.py meeting.mp4 --diarize --hf-token $HF_TOKEN

Output becomes:

**[SPEAKER_00]** 00:00 Let's start today's review.
**[SPEAKER_01]** 00:08 Sounds good, I'm ready.

(HuggingFace token for pyannote diarization model is free to get — just click "Agree" on the model page.)

🏗️ Architecture

video_transcription/
├── video_transcribe.py       # CLI orchestrator
├── src/
│   ├── audio.py              # mlx_whisper wrapper + video→wav extract
│   ├── frames.py             # ffmpeg frame extraction (3 modes + dedup)
│   ├── merge.py              # timeline builder + md/json renderers
│   ├── preflight.py          # dependency check (--check)
│   └── agents/
│       ├── base.py           # BaseAgent(ABC) + AgentInvocationError
│       ├── claude.py         # claude -p + Read-tool image pattern
│       ├── codex.py          # codex exec -i
│       └── gemini.py         # gemini -p @path
├── prompts/
│   └── default.md            # built-in frame-description prompt
├── examples/
│   ├── meeting_recipe.md
│   └── lecture_recipe.md
└── tests/                    # 31 unit tests, all pure-function where possible

Design principles:

• 🎯 One responsibility per file. frames.py only knows frames. merge.py only knows timelines. Adapters only know their one CLI.
• 🧪 TDD on pure functions. The hybrid frame-selection algorithm (min/max interval + downsample + dedup) has 10 unit tests and zero subprocess dependencies.
• 🚫 No shell=True. Everything is a list[str] to subprocess.run. Zero injection surface.
• 🔁 Retry-once on CLI failures. Transient network errors don't kill the whole run — just that one frame gets a [frame description failed: ...] placeholder and the pipeline continues.

🧪 Dev

python3 -m pytest -v    # 31 tests

📚 Built On

• mlx-whisper — Metal-accelerated Whisper
• whispermlx — diarization via pyannote
• ffmpeg — audio/frame extraction
• treesoop/whisper_transcription — the audio pipeline started here, got forked and expanded

🤔 Comparison

	video_transcription	paid SaaS (Notta, etc.)	whisper + custom scripts
Audio transcript	✅	✅	✅
Frame descriptions	✅	❌	❌ (you'd build it)
Context-aware frames	✅ (audio → agent)	—	—
Speaker diarization	✅ (optional)	✅	⚠️ (setup pain)
Runs locally	✅	❌	✅
Hallucination guarded	✅	✅	⚠️ (default settings are lossy)
Multilingual	✅ 99 langs	✅	✅
Price	Free	Subscription	Free (but your time)
Swap AI provider	✅ (one flag)	❌	—

📜 License

MIT — same as upstream. Do whatever you want, attribution appreciated.

🙌 Contributing

PRs welcome. Good places to start:

• Add a new agent adapter (Aider, Cursor headless, local LM Studio, etc.) — look at src/agents/gemini.py for the ~30-line template
• Imagehash-based visual dedup (current MD5 dedup only catches bitwise-identical PNGs; visually-similar frames still get described multiple times)
• Single-pass ffmpeg select filter for _extract_at_timestamps (current: one ffmpeg call per frame, fine up to ~100 frames)
• Windows / Linux support (currently macOS-only because of mlx_whisper)

---

Made by Dion · Part of Treesoop

If this is useful, please ⭐ the repo — it's how I know what's worth building next.