whisper-vtt2srt

A robust, production-grade library designed to convert WebVTT to SRT, turning messy AI transcripts into clean, usable subtitles.

A post-processing tool designed to clean the output from OpenAI Whisper, YouTube Auto-Captions, and other AI transcription services.
Perfect for TTS pipelines, video dubbing, and dataset preparation.

🧠 The Problem with Raw AI Subtitles

AI tools like Whisper are incredible at speech recognition, but their raw VTT output is often chaotic. They frequently produce:

• The "Karaoke Effect": Words accumulating screen-by-screen (e.g., "Hello", "Hello world", "Hello world!").
• Micro-Glitches: Subtitle frames lasting milliseconds that are invisible to humans but break TTS/dubbing scripts.
• Metadata Clutter: Tags like align:start, <c>, <b> or <i> that mess up text processing.

whisper-vtt2srt is the bridge between raw AI output and your production pipeline. It stabilizes and normalizes the text, making it safe for Text-to-Speech (TTS) generation, video players, and NLP tasks.

🚀 Try it Online

Test the conversion instantly in your browser (Client-Side / Secure). No installation required.

📖 Table of Contents

• 🎮 Online Playground
• 👀 See the Difference
• 🚀 Key Features
• 📦 Installation
• 📘 How to Use
  • 💻 CLI Usage
    • Command Help
  • 🐍 Python API Usage
    • Basic Conversion
    • Advanced Control
• 🧠 How It Works
• 📆 Changelog
• 🤝 Contributing
• 📜 License
• 📚 Reference

👀 See the Difference (Before vs After)

WEBVTT
Kind: captions
Language: en

00:00:00.640 --> 00:00:03.110 align:start position:0%
 
APIs<00:00:01.280><c> are</c><00:00:01.520><c> everywhere.</c><00:00:02.399><c> They</c><00:00:02.639><c> power</c><00:00:02.960><c> your</c>

00:00:03.110 --> 00:00:03.120 align:start position:0%
APIs are everywhere. They power your
 

00:00:03.120 --> 00:00:05.430 align:start position:0%
APIs are everywhere. They power your
apps,<00:00:03.600><c> your</c><00:00:03.840><c> payment</c><00:00:04.160><c> systems,</c><00:00:04.880><c> your</c><00:00:05.120><c> cloud</c>

00:00:05.430 --> 00:00:05.440 align:start position:0%
apps, your payment systems, your cloud
 

00:00:05.440 --> 00:00:07.829 align:start position:0%
apps, your payment systems, your cloud
services,<00:00:06.560><c> pretty</c><00:00:06.879><c> much</c><00:00:07.120><c> every</c><00:00:07.440><c> piece</c><00:00:07.680><c> of</c>

00:00:07.829 --> 00:00:07.839 align:start position:0%
services, pretty much every piece of
 

00:00:07.839 --> 00:00:10.470 align:start position:0%
services, pretty much every piece of

1
00:00:00,640 --> 00:00:03,110
APIs are everywhere. They power your

2
00:00:03,120 --> 00:00:05,430
apps, your payment systems, your cloud

3
00:00:05,440 --> 00:00:07,829
services, pretty much every piece of

🚀 Key Features

• 🛡️ Stabilization Strategy Intelligently detects and merges accumulating text blocks ("Karaoke Effect"), preventing the rapid flashing of partial sentences. Essential for generating smooth audio in TTS pipelines, video dubbing, and subtitles.

• 🎵 Sound Description Removal Automatically filters out non-speech elements like [Music], [Applause], or [Laughter], ensuring your TTS voice doesn't try to read stage directions.

• 🧹 Glitch Filtering Automatically removes subtitle blocks with insignificant duration (< 50ms) that can cause audio generation errors or player flickering.

• ✨ Smart Normalization Strips VTT-specific metadata (align:start, position:0%), removes internal tags (<c>, <00:00:00>), and cleans up inconsistent whitespace ensuring pure text output.

• ⚡ Zero Dependencies Built with pure Python standard library. Lightweight and easy to install in any environment (Linux, Windows, Docker).

• 🔧 Configurable Strictness Every cleaning step is optional. You enable exactly what your pipeline needs.

📦 Installation

pip install whisper-vtt2srt

📘 How to Use

💻 CLI Usage

Process files directly from the command line:

# Convert a Single File
whisper-vtt2srt input.vtt

# Batch Convert a Folder
whisper-vtt2srt ./my_dataset

# Recursive Conversion (subfolders included)
whisper-vtt2srt ./my_dataset --recursive

# Handle Legacy Encodings (e.g., Latin-1)
whisper-vtt2srt input_latin.vtt --encoding ISO-8859-1

# Keep the "karaoke" effect (disable deduplication)
whisper-vtt2srt input.vtt --no-karaoke

Command Help

usage: whisper-vtt2srt [-h] [-r] [-e ENCODING] [--no-karaoke] [--keep-glitches] [--keep-formatting]
               [--keep-metadata] [--merge-short-lines]
               input [output]

Convert WebVTT to SRT with professional cleaning.

positional arguments:
  input                 Input VTT file or directory
  output                Output SRT file or directory (optional)

options:
  -h, --help            show this help message and exit
  -r, --recursive       Recursively process directories
  -e ENCODING, --encoding ENCODING
                        Input file encoding (default: utf-8)
  --no-karaoke          Disable anti-karaoke filter (keep accumulating text)
  --keep-sound-descriptions
                        Keep sound descriptions like [Music] or [Applause]
  --keep-glitches       Keep short <50ms blocks
  --keep-formatting     Keep VTT tags (bold, italic, colors)
  --keep-metadata       Keep metadata tags (align:start, position:0%)
  --merge-short-lines   Aggressively merge short lines into single lines
  --max-line-length MAX_LINE_LENGTH
                        Maximum line length allowed when merging short lines (default: 42, like YouTube/Netflix)

🐍 Python API Usage

Easily integrate whisper-vtt2srt into your own Python pipelines. The library exports a high-level Pipeline class for full control.

Basic Conversion

from whisper_vtt2srt import Pipeline

# 1. Initialize
pipeline = Pipeline()

# 2. Read input
with open("subs.vtt", "r", encoding="utf-8") as f:
    raw_vtt_content = f.read()

# 3. Convert raw VTT content
srt_content = pipeline.convert(raw_vtt_content)

# 4. Use the clean SRT (e.g., send to TTS engine, save to file, render in player, etc.)
print(srt_content)

Advanced Control

You can customize the cleaning options if needed:

Just pass a CleaningOptions object to the Pipeline constructor to toggle specific cleaning rules.

from whisper_vtt2srt import CleaningOptions, Pipeline

# Configure strictness
options = CleaningOptions(
    remove_pixelation=True,    # Fix Karaoke effect
    remove_sound_descriptions=True, # Remove [Music], [Applause]
    remove_glitches=True,      # Remove <50ms blocks
    simplify_formatting=True,  # Strip tags like <c> or <b> and fix whitespace
    remove_metadata=True,      # Clean VTT positioning
    merge_short_lines=False,   # Aggressively merge short lines
    max_line_length=42         # Max length constraint for merging
)

pipeline = Pipeline(options)

🧠 How It Works

1. Parser (State Machine): Robustly reads messy VTT files, handling multi-line strings and irregular spacing.
2. Deduplication Engine: Uses a sliding window to identify comparison patterns between blocks. If a block is just a prefix of the next one (common in AI streams), it is merged or removed to stabilize the text.
3. Filter Layer: Applies duration checks and regex cleaning to ensure the final output is compliant with the SubRip (SRT) standard.

📆 Changelog

Project history and updates are tracked in CHANGELOG.md.

🤝 Contributing

Contributions are welcome! We follow a strict SOLID architecture. See CONTRIBUTING.md for details.

📜 License

MIT License - see LICENSE.

📚 Reference

If you use this library in your research or project, please cite it as:

@software{whisper_vtt2srt,
  author = {Jorcelino Junior},
  title = {whisper-vtt2srt: A robust WebVTT to SRT converter for AI subtitles},
  year = {2026},
  url = {https://github.com/jorcelinojunior/whisper-vtt2srt}
}

Saved you time? Helped your project?

Support independent open-source development!