whispersub 1.0.2

Transcribe videos to ASS subtitle files using faster-whisper

whispersub

Transcribe video files to ASS subtitle files using faster-whisper (Whisper large-v3-turbo).

• Batch mode: pass multiple files or directories to scan recursively
• Surround-sound audio extraction (dialogue-channel aware)
• Word-level timestamps with balanced line breaking
• Per-word confidence colour coding in the terminal

Requirements

Python 3.10+. Works on Linux, Windows, and macOS. No system FFmpeg needed — PyAV bundles its own.

Install

pip install whispersub

GPU acceleration (Linux/Windows — requires an NVIDIA GPU with CUDA 12):

pip install whispersub[gpu]

Without [gpu], whispersub falls back to CPU automatically if CUDA is unavailable.

Usage

# Single file — writes movie.en.ass alongside the video
whispersub movie.mkv

# Explicit output directory
whispersub movie.mkv --output-dir ~/subs

# Whole directory, force overwrite
whispersub /media/shows --force

# File with multiple audio tracks — list tracks first, then pick one
whispersub series.mkv                     # error lists available tracks
whispersub series.mkv --audio-track 2

Options

Option	Default	Description
--audio-track N	auto	Audio track index (required if the file has multiple tracks)
--colour-by	probability	Per-word terminal colour coding: probability or duration
--font-size N	48	Font size (1280×720 canvas; player scales to actual resolution)
--force	off	Overwrite existing subtitle files
--limit N	—	Stop after N segments per video (useful for testing)
--max-line-count N	2	Maximum subtitle lines per card
--max-line-width N	36	Maximum characters per line
--max-threads N	all cores	CPU thread limit
--output-dir DIR	alongside video	Write all subtitle files to this directory

Output

Subtitle files are named <stem>.<language>.ass, e.g. movie.en.ass. The detected language comes from Whisper.

Each ASS file contains hidden Comment: events carrying word-level timestamp data in JSON, so the raw timing can be recovered by tooling without re-transcribing.

Licence

MIT