Anysub (formerly Whishper) makes subtitles from audio and video. Upload a recording, let Whisper transcribe it, then fix the text and timing in the editor and export. It runs on your own computer or server.
- Transcription with Whisper, from the
tinymodel tolarge-v3-turbo. - Word-level timings.
- A subtitle editor with a waveform timeline, undo and autosave.
- Warnings for subtitles that are too fast, too long, or overlap.
- Export to SRT, WebVTT, plain text or JSON.
- Uploads that resume, for hours-long files.
- A searchable library.
- Admin pages for models, defaults, upload limits, disk space and workers.
You need Linux on x86-64 with Docker and Docker Compose.
-
Create a folder and download the Compose file:
mkdir anysub && cd anysub curl -fsSLO https://github.com/pluja/anysub/releases/latest/download/compose.yaml
-
Optional. Skip to step 3 to transcribe on the CPU. Otherwise, download one extra file into the same folder and turn it on:
-
NVIDIA GPU. You need the NVIDIA Container Toolkit and a GTX 16xx, RTX 20xx or newer card.
curl -fsSLO https://github.com/pluja/anysub/releases/latest/download/compose.cuda.yaml echo "COMPOSE_FILE=compose.yaml:compose.cuda.yaml" >> .env
-
External providers only. If you only transcribe with a service such as OpenAI or Groq, this smaller worker leaves out the local models and needs no GPU. Add the service under Administration > External providers once Anysub runs. See External providers.
curl -fsSLO https://github.com/pluja/anysub/releases/latest/download/compose.remote.yaml echo "COMPOSE_FILE=compose.yaml:compose.remote.yaml" >> .env
-
-
Start Anysub:
docker compose up -d
-
Open
http://localhost:5173and create your account. It asks for a setup code, whichdocker compose logs webshows.
A GPU transcribes several times faster than a CPU. AMD and Intel GPUs aren't supported yet, but we plan to add them.
Install covers Podman, reverse proxies and building from source. NVIDIA GPUs helps when the GPU doesn't work.
Whishper 3 is no longer maintained, and Anysub can't read its data. Coming from Whishper 3 explains how to move over. The old code is in the v3-legacy branch.
- Translation.
- Speaker labels.
- Watched folders.
- Import from a URL or an existing subtitle file.
- More than one account.
- Workers on other machines.
- ARM images.
These will come in 4.x releases. See the changelog.
See docs.
Large language models helped write the code and docs. A human reviews and tests every change before it ships.
