Anime Transcribe

Warning

It only supports Mac as a deliberate design choice. This project is vibe coded, so the minimalistic scope is required to make sure AI models can actually make something decent.

Utilizes silero VAD for detecting when charaters are speaking, then uses whisper-ja-anime-v0.1 for generating "native" subtitles.

Features:

• Supports GPU acceleration trough MPS.
• Offline first. The models will be downloaded and cached in the ~/.cache folder (~2GB-4GB in total). The script will then use those, instead of pinging the network.

Usage

Prerequisites

ffmpeg needs to be installed.

Install

git clone https://github.com/mintc2/anime-transcribe
cd anime-transcribe
uv sync

Use

Transcribe with the default output path:

uv run anime-transcribe /path/to/video/file

Transcribe to a user defined output dir:

anime-transcribe -o /path/where/to/output /path/to/video/file

During the first run, the script will downloaded the weights of the used models (~2GB). This might take a while depending on the internet connection quality.

Uninstall

Delete the cloned project

rm -rf /wherverer/its/installed/anime-transcribe

Clean the bulky uv cache and the installed models

rm -rfi ~/.cache/uv ~/.cache/huggingface ~/.cache/torch