NOTE: This is a vibe coded app. Proceed with caution.

add_subs_to_videos

You have video files but no subtitles. This tool fixes that.

It crawls a directory recursively, transcribes every video it finds using whisper.cpp, and writes a .srt sidecar file next to each one. Already-transcribed files are skipped, so re-running is fast.

Repo

https://github.com/simnim/add_subs_to_videos

Install

Every option below gives you both the add_subs_to_videos CLI command and an "Add Subs to Videos" desktop GUI.

Ubuntu (snap — recommended, no prerequisites):

sudo snap install add-subs-to-videos

Installs the add-subs-to-videos CLI plus an "Add Subs to Videos" launcher in your applications menu (add-subs-to-videos-gui).

macOS (.dmg — no prerequisites): Download Add Subs to Videos-<version>.dmg from GitHub Releases, open it, and drag the app to Applications. The DMG also includes a standalone add_subs_to_videos CLI binary and an Install CLI.command helper — double-click it to symlink the CLI into /usr/local/bin so add_subs_to_videos --help works from a terminal.

Metal GPU acceleration is auto-detected on macOS — no extra steps needed.

Any platform with Python 3.12+ (via pip/pipx):

pip install "add-subs-to-videos[gui]"
# or
pipx install "add-subs-to-videos[gui]"

This installs both the add_subs_to_videos CLI and the add-subs-to-videos-gui GUI entry point. Drop [gui] if you only want the CLI.

pywhispercpp compiles a C++ extension at install time (requires CMake — brew install cmake ffmpeg on macOS, or on Ubuntu/Debian sudo apt install cmake ffmpeg). ffmpeg/ffprobe are also needed at runtime to decode non-WAV media and show per-file progress. On Linux with an NVIDIA GPU, set WHISPER_CUDA=1 to compile with CUDA support instead of falling back to CPU:

WHISPER_CUDA=1 pip install "add-subs-to-videos[gui]" --no-cache-dir

Requires the CUDA Toolkit (nvcc on PATH) already installed. --no-cache-dir forces a rebuild — without it, pip may reuse a cached non-CUDA wheel from a previous install. If you already installed without WHISPER_CUDA=1, reinstall with --force-reinstall --no-cache-dir to pick up the flag.

Usage

CLI

add_subs_to_videos /path/to/videos
add_subs_to_videos /path/to/videos --model large-v3 --language en
add_subs_to_videos /path/to/videos --force   # re-transcribe even if .srt exists
add_subs_to_videos /path/to/videos --threads 4   # default: all available CPU cores
add_subs_to_videos /path/to/videos --quiet   # only warnings/errors + final summary
add_subs_to_videos /path/to/videos --verbose # debug output, incl. detected language

movie.mp4 → movie.srt, placed in the same directory. Supports .mp4, .mkv, .avi, .mov, .m4v, .webm, .ts, .flv.

Your last-used directory, model, language, and threads are remembered in ~/.config/add-subs-to-videos/config.toml, shared between the CLI and GUI — e.g. pick a folder in the GUI and a later CLI run can omit the directory argument.

GUI

Drag a folder (or single video) onto the window, or click to pick one, then hit Run. Includes handy progress bars for model downloads and live, per-file transcription logs.

• A file table lists every discovered video with a live status (Pending / Processing / Done / Skipped / Failed) and a clickable log icon per row that opens that file's transcript/log output.
• Model dropdown (tiny/base/small/medium/large-v3) shows whether the model is already downloaded, not yet downloaded, or currently downloading.
• Language dropdown offers auto-detect or any of whisper.cpp's supported languages.
• A threads spinbox, a "Force re-run" checkbox, and a "Debug logging" checkbox.
• Overall and per-file progress bars, plus a Cancel button that aborts the current download or in-progress file.
• Auto re-run: once a run finishes (and wasn't cancelled), the Run button relabels to a countdown — "Run (auto re-run in M:SS)" — and automatically rescans the folder and re-runs after 10 minutes, so it can be left running unattended to pick up new files as they appear. Clicking Cancel during the countdown stops it.

Partial subtitles & cancellation

While a file is transcribing, segments are streamed incrementally into a movie.srt.part sidecar next to the eventual movie.srt, so you can open a growing, valid SRT file mid-run. On success it's atomically renamed to movie.srt. If you cancel (or the process is interrupted), the .part file is left on disk as-is — it's never deleted, but it's also never resumed: a later run on a file that only has a stray .part (no finished .srt) re-transcribes it from scratch, overwriting the partial content. Files that already have a completed .srt are still skipped unless --force/"Force re-run" is set.

Models

medium is the default. Use large-v3 when accuracy matters; use tiny or base for quick drafts on long files.

Output

Each video gets a .srt file in the same directory:

movie.mp4  →  movie.srt

1
00:00:01,240 --> 00:00:04,120
Welcome to the presentation.

2
00:00:04,500 --> 00:00:07,300
Today we'll cover three topics.