subtitle-cleaner 0.3.3

A local-first subtitle sanitizer, readability optimizer, and NLE import formatter

subtitle-cleaner

A local-first Python CLI toolkit for sanitizing, repairing, and optimizing AI-generated subtitle files.

AI transcription tools (Whisper, CapCut, Premiere's auto-transcribe) save time up front — but the cleanup afterward is brutal. Stuttered words, broken import syntax, filler noise, and ugly line breaks pile up fast.

subtitle-cleaner runs entirely on your machine with no cloud dependencies, no privacy risk, and no subscription.

---

✨ Features

Feature	Flag	Description
Duplicate word removal	(default)	Case-insensitively removes stutters ("We're we're" → "We're")
Filler word stripping	(default)	Removes um, uh, like without breaking timings
HTML/font tag stripping	(default)	Cleans embedded <font>, <i>, <b> tags from SRT/VTT
Arrow & timestamp repair	(default)	Normalizes unicode arrows (–>, ->) to standard -->
Index re-sequencing	(default)	Fixes out-of-order subtitle block indices starting from 1
Timestamp normalization	(default)	Pads and corrects malformed HH:MM:SS,mmm timestamps
Repair preview (diff)	--preview / -p	Non-destructive diff of proposed changes before saving
Semantic line breaking	--segment / -s	Splits long lines at grammatical boundaries (not character counts)
Mobile formatting	--mobile / -m	30-character line width for 9:16 vertical video (TikTok/Shorts/Reels)
Custom vocabulary map	--vocab vocab.json	JSON-based find-and-replace for brand names, jargon, speaker names
NLE optimization	--nle premiere|resolve	Premiere Pro (37-char limit, 2-line max) and Resolve (UTF-8 BOM) modes
Batch processing	-i /folder/ -o /out/	Recursively processes entire directories of .srt/.vtt files
Format conversion	--format srt|vtt|ass / -f srt|vtt|ass	Convert between SRT, VTT, and ASS formats (converts timestamps & styles)
Word-level splitting	--word-split / -w	Splits blocks into single-word timed subtitle blocks (proportionally distributed duration)
Karaoke style export	--karaoke / -k	Exports to ASS format with highlighting {\k} centisecond timings
YouTube caption sync	youtube_sync.py	Downloads, cleans, and saves captions from any YouTube URL

---

🚀 Installation

# Install directly from GitHub
pip install git+https://github.com/r1ngotchi/subtitle-cleaner.git

# Or clone and install in editable mode for development
git clone https://github.com/r1ngotchi/subtitle-cleaner
cd subtitle-cleaner
pip install -e .

💻 Quick Start

Once installed, the toolkit provides three convenient CLI commands:

# Preview what would change (non-destructive)
subtitle-cleaner -i messy.srt --preview

# Clean and save output
subtitle-cleaner -i messy.srt -o clean.srt

# Clean for Premiere Pro import (37-char line limit, 2-line max)
subtitle-cleaner -i messy.srt -o clean.srt --nle premiere

# Clean for DaVinci Resolve (UTF-8 BOM encoding)
subtitle-cleaner -i messy.srt -o clean.srt --nle resolve

# Semantic line breaking + mobile formatting
subtitle-cleaner -i messy.srt -o clean.srt --segment --mobile

# Apply custom vocabulary corrections
subtitle-cleaner -i messy.srt -o clean.srt --vocab my_vocab.json

# Batch process an entire folder
subtitle-cleaner -i ./subtitles/ -o ./cleaned/

# Convert SRT to VTT (strips indices, normalizes dots)
subtitle-cleaner -i input.srt -o output.vtt -f vtt

# Convert VTT to SRT (restores sequential indices, normalizes commas)
subtitle-cleaner -i input.vtt -o output.srt -f srt

# Split subtitles into single-word blocks (e.g. for vertical video captions)
subtitle-cleaner -i input.srt -o output.srt -w

# Convert to ASS format with karaoke highlighting (for animated captions)
subtitle-cleaner -i input.srt -o output.ass -k

# Download and clean YouTube captions directly
subtitle-youtube-sync https://www.youtube.com/watch?v=VIDEO_ID -o output.vtt

---

📋 Before / After Example

Input (messy.srt):

1
00:00:01,000 -> 00:00:04,000
Yeah, we're we're going to like, um, build this.

1
00:00:04,100 ---> 00:00:06,000
Like, uh, <font color="#ff0000">absolutely.</font>

Output after python cleaner.py -i messy.srt:

1
00:00:01,000 --> 00:00:04,000
Yeah, we're going to build this.

2
00:00:04,100 --> 00:00:06,000
Absolutely.

---

📦 Custom Vocabulary Map (--vocab)

Create a vocab.json file mapping AI mistranscriptions to their correct forms:

{
  "open eye": "OpenAI",
  "adobe premiere pro": "Adobe Premiere Pro",
  "da vinci resolve": "DaVinci Resolve"
}

Then run: python cleaner.py -i messy.srt -o clean.srt --vocab vocab.json

---

🔬 Diagnostics

Run the linter to get a full report of issues before cleaning:

subtitle-diagnostics messy.srt

The linter checks for:

• Reading speed — flags blocks with dangerously high CPS (characters per second)
• NLE compatibility — warns about 3+ line blocks and lines >37 chars (Premiere crash risk)
• Timing overlaps — detects blocks where end time > next block's start time
• Whitespace corruption — tabs, trailing spaces, CRLF issues

---

⚙️ YouTube Caption Sync

subtitle-youtube-sync https://www.youtube.com/watch?v=VIDEO_ID -o captions.vtt

Downloads auto-generated or manually uploaded captions, cleans them, and saves a polished file ready for upload or NLE import.

---

🧪 Tests

Unit Tests

Verify individual module features and parser stability:

python -m unittest test_cleaner.py
# Expected: 19 tests, all passing

Regression Tests

Test the cleaner automatically against all collected real-world corruptions in the dataset:

subtitle-regression
# Expected: Runs and passes all dataset test cases with 0 critical errors remaining

---

📁 Project Structure

subtitle-cleaner/
├── cleaner.py          # Core CLI tool
├── diagnostics.py      # Linting & validation engine
├── youtube_sync.py     # YouTube caption downloader-cleaner
├── detectors/          # Modular lint checkers
│   ├── reading_speed.py
│   ├── nle_compatibility.py
│   └── ...
├── sample_input.srt    # Test fixture
├── sample_vocab.json   # Example vocabulary map
└── test_cleaner.py     # 15 unit + integration tests

---

🤝 Contributing

Found a subtitle file that breaks the parser? Submit an issue and attach the file — we're building a structured corruption dataset to improve repair reliability.

Pull requests welcome. See CONTRIBUTING.md for guidelines.

---

☕ Support & Funding

If subtitle-cleaner saves you hours of manual editing, consider supporting the project:

• Cryptocurrency: Support us with BTC, ETH, SOL, or Polygon in DONATIONS.md
• GitHub Sponsors: Sponsor r1ngotchi

---

📄 License

MIT — free to use, modify, and distribute.