vegF1 - Formula 1 Video Profanity Replacement

Automatically replaces profanity in F1 videos with vegetable names using AI speech synthesis.

Origin Story

This project was inspired by a comment in the Missed Apex podcast:

"Surely AI can replace the words with random fruits and vegetables"

— Missed Apex Podcast, 26:02

This project takes a random comment and thru the power of laziness and delayed project time makes it real: automatically replace profanity with vegetable names using AI.

The solution: replace every swear word with a random vegetable name, spoken in a natural voice. "What the fuck?" becomes "What the butternut squash?" The result is absurd and family-friendly?

It's terrible, but it's our terrible.

Examples

Try it yourself with these F1 team radio clips:

# Example 1: Kevin Magnussen door incident (22 replacements)
uv run python dl-video.py "https://www.youtube.com/watch?v=tnY65NRvwUQ"

# Example 2: Classic team radio moments
uv run python dl-video.py "https://www.youtube.com/watch?v=5h4FDy9Eqzs"

Example Replacements:

• "fucking blind asshole" → "broccoli... blind... kale" (preserves non-swears!)
• "fucking motherfucker" → "bamboo shoots, fenugreek, chard"
• "fucking dickhead" → "collard greens, garlic"
• "fuck sake" → "rhubarb sake" (preserves "sake"!)

Quick Start

# Download and process a video
uv run python dl-video.py [VIDEO_URL]

# The script will:
# 1. Download video and extract audio
# 2. Transcribe with WhisperX (word-level timing)
# 3. Replace profanity with vegetable names (TTS)
# 4. Merge edited audio back with video
# 5. Output: final_output.mp4

Project Structure

vegF1/
├── dl-video.py              # Main entry point
├── timeline_processor.py    # Core audio processing library
├── word_lists.json          # Swear words and vegetable names
├── final_output.mp4         # Latest generated video
│
├── scripts/                 # Utility scripts
│   ├── compare.py          # Unified comparison tool
│   ├── verify.py           # Verification tool (transcribe + check)
│   ├── create_video.py     # Video creation tool
│   └── test_*.py           # Test scripts
│
├── outputs/                 # All generated files
│   ├── *.wav               # Audio files
│   ├── *.png               # Visualizations
│   ├── *.log               # Replacement logs
│   └── *.txt               # Transcripts
│
├── docs/                    # Documentation
│   ├── REFACTOR_SUMMARY.md
│   └── improvements_summary.md
│
├── archive/                 # Old/deprecated files
│   └── audio_processor.py  # Original processor
│
├── audio/                   # Downloaded audio files
├── videos/                  # Downloaded video files
└── tts_cache/              # Cached TTS audio

Consolidated Tools

1. Compare Audio/Visualizations

Replaces: create_aligned_comparison.py, create_proper_comparison.py, create_wordlevel_comparison.py, enhanced_waveform_comparison.py

cd scripts

# Full comparison
uv run python compare.py \
  --original ../audio/original.wav \
  --edited ../outputs/edited_audio.wav \
  --output ../outputs/comparison.png

# Zoomed comparison (specific time range)
uv run python compare.py \
  --original ../audio/original.wav \
  --edited ../outputs/edited_audio.wav \
  --output ../outputs/comparison_zoomed.png \
  --zoom-start 4.5 \
  --zoom-end 8.0

# With transcript comparison
uv run python compare.py \
  --original ../audio/original.wav \
  --edited ../outputs/edited_audio.wav \
  --transcript-original ../outputs/transcript.txt \
  --transcript-edited ../outputs/edited_transcript.txt \
  --transcript-output ../outputs/transcript_comparison.txt

2. Verify Output

Replaces: verify_with_transcription.py, check_edited.py

cd scripts

# Verify edited audio has no swear words
uv run python verify.py \
  --audio ../outputs/edited_audio.wav \
  --swear-words ../word_lists.json \
  --output ../outputs/verification_transcript.txt

# Exit code 0 = success (no swears found)
# Exit code 1 = failure (swears found)

3. Create Video

Replaces: create_final_video.py

cd scripts

# Create video from processed audio
uv run python create_video.py \
  --video ../videos/video.mp4 \
  --audio ../audio/audio.wav \
  --segments segments.json \
  --output ../outputs/final_output.mp4 \
  --debug

How It Works

Timeline-Based Processing

The core of this project is a declarative timeline-based architecture that:

1. Phrase Detection: Groups swear words within 2000ms into phrases

   • Example: "Fucking hell" with 1.5s gap → single phrase replacement

2. Word-Level Replacement: Only replaces profanity + 400ms buffer

   • "Kevin, just fucking smash the door" → "Kevin, just [vegetables] smash the door"
   • NOT: "[vegetables]" (which segment-level would do)

3. Multiple Vegetables: Fills longer phrases with multiple vegetables

   • "Fucking hell" (2.4s) → "butternut squash, broccolini, turnip" (2.3s)
   • Natural timing with 95%+ duration matching

4. Punctuation & Variants: Handles "hell." vs "hell", "fuckin" vs "fucking"

Processing Pipeline

Input Video
    ↓
Extract Audio (WAV)
    ↓
Transcribe with WhisperX (word-level timing)
    ↓
Create Timeline (keep/replace segments)
    ↓
Generate TTS for vegetables
    ↓
Build edited audio from timeline
    ↓
Merge audio + video with ffmpeg
    ↓
Output Video

Configuration

word_lists.json

{
  "swear_words": ["fuck", "shit", "hell", ...],
  "vegetable_names": ["bamboo shoots", "brussels sprouts", ...]
}

Timeline Processor Settings

See timeline_processor.py:

• safety_buffer_ms: 400ms (buffer after each swear)
• phrase_gap_threshold: 2000ms (max gap to group swears)
• stretch_limit: 0.7-1.5x (TTS speed adjustment range)

Development

Running Tests

cd scripts
uv run python test_timeline_with_words.py

Adding New Vegetables

Edit word_lists.json and add to vegetable_names array.

Debugging

1. Enable debug mode: --debug flag
2. Check outputs in outputs/ folder
3. Review replacement log: outputs/*_replacements.log
4. Verify with: scripts/verify.py --audio outputs/edited_audio.wav

Results

• 10+ replacements per typical F1 video
• 95%+ duration matching (sounds natural)
• 0 swear words in output (verified by transcription)
• 22% time saved (removed profanity + buffers)

License

MIT