█

AI/TLDR

Buy me a coffee
Buy me a coffee

Anysub

A self-hosted subtitle studio (formerly Whishper) that transcribes audio and video on your own machine

Audio, Music & VoiceOpen source
Latest
v4.0.0
Updated
10 Nov 2026
Language
Svelte
License
AGPL-3.0

What's new

v4.0.010 Nov 2026

Whishper was rewritten from scratch and renamed Anysub: WhisperX transcription with word timings, a waveform subtitle studio and a two-container SQLite stack. It cannot read Whishper 3 data, and translation and URL imports are not in 4.0.

IN SHORT

What is Anysub?

Anysub is an open-source web app that turns audio and video into subtitles, with everything running on your own machine. It sits in the Audio, Music & Voice category of the Multimodal Generation tools and is open-source, written in Svelte and licensed AGPL-3.0. To get started: download the Compose file and start it.

Updated Markdown version ↗

Overview

Anysub is an open-source web app that turns audio and video into subtitles, with everything running on your own machine. You upload a recording, it transcribes the speech with Whisper models from `tiny` to `large-v3-turbo`, and you correct the result in a browser subtitle studio before exporting SRT, WebVTT, plain text or JSON. It is the successor to Whishper by the same maintainer, rewritten from scratch and published under the new name with version 4.0.0 on November 10, 2026; the GitHub repository moved from `pluja/whishper` to `pluja/anysub`.

Transcription runs in a Python worker built on WhisperX, which uses Faster-Whisper and CTranslate2 for speech recognition and forced alignment for word-level timings. Anysub checks the worker's hardware and marks each model as recommended, runs well, may be slow or not recommended, and downloads a model the first time you use it. After that, transcription works offline: the project states that it sends no telemetry and never writes transcript text into its logs.

The studio is where most of the time goes. The video sits beside the cue list, a waveform timeline runs underneath, and you can split, merge and retime cues by dragging or with the keyboard, with undo, redo and autosave. Each cue shows its duration and characters-per-second rate, and Anysub warns about cues that are too fast, too short, too long or overlapping. Uploads resume after a dropped connection, so hours-long recordings are practical, and a searchable library can match subtitle text as well as project names.

The Anysub studio: a video of the ISS crew plays on the left with a burned-in subtitle, the right-hand pane lists numbered cues with start and end timestamps, durations and a quality-warnings count, and a waveform timeline with cue blocks runs along the bottom
The studio puts the media beside the cue list, with a waveform timeline underneath for dragging cues and a quality-warning count above the list.Anysub README ↗

The rewrite is deliberately smaller than Whishper 3: two containers instead of a stack with MongoDB, LibreTranslate and Nginx, an SQLite database, and no configuration before the first start. It cannot read Whishper 3 data, and some Whishper 3 features are not in the 4.0 line yet — translation, transcribing from a URL, importing subtitle files, multiple accounts and ARM images are listed as planned for later 4.x releases. Anysub is AGPL-3.0-only.

What it does

  • Whisper transcription through WhisperX, from `tiny` to `large-v3` and `large-v3-turbo`, in every language Whisper supports, with hardware-aware model recommendations
  • Word-level timings from forced alignment, kept in the JSON export
  • A subtitle studio with a waveform timeline, synced playback, split, merge, drag or keyboard retiming, undo/redo and autosave with crash recovery
  • Quality warnings for cues that are too fast, too short, too long or overlapping, with a per-cue characters-per-second reading
  • Export to SRT, WebVTT, plain text and Anysub JSON
  • Resumable uploads, a searchable project library, and admin pages for models, transcription defaults, upload limits, disk space and workers

Getting started

Anysub runs as two containers — the web app and a transcription worker — from one Compose file, on Linux x86-64 with Docker and the Compose plugin (or Podman). There is no `.env` to fill in before the first start.

Download the Compose file and start it

The first start pulls the web and worker images and sets everything up. Plan on about 6 GB of disk for the images and a first model, and 4 GB of free memory for the `Small` model.

bashbash
mkdir anysub && cd anysub
curl -fsSLO https://github.com/pluja/anysub/releases/latest/download/compose.yaml
docker compose up -d

Create the owner account

Open `http://localhost:5173`. Anysub asks for a setup code, which it prints in the web container's log so only someone with access to the server can claim the account. This is the only account; there is no sign-up.

bashbash
docker compose logs web

Optional: switch the worker to an NVIDIA GPU

A GPU transcribes several times faster. It needs a GTX 16xx / RTX 20xx or newer, the NVIDIA Container Toolkit and a driver that supports CUDA 12.8, and Docker rather than Podman. Download the GPU Compose file next to `compose.yaml` and point Compose at both.

bashbash
curl -fsSLO https://github.com/pluja/anysub/releases/latest/download/compose.cuda.yaml
echo 'COMPOSE_FILE=compose.yaml:compose.cuda.yaml' >> .env
docker compose up -d

Transcribe, edit and export

Create a project by uploading a file in the library, pick a model (`Small` is the suggested start on a CPU), leave word alignment on, choose the language or let it detect, and start the transcription. The studio opens with the video, the cue list and the waveform; fix the wording and timings, then export SRT, WebVTT, text or JSON.

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

When to use it

  • Reach for it when the recording is confidential — an interview, a medical or legal conversation, an internal meeting — and a hosted transcription API is not an option
  • Reach for it to subtitle a video end to end: transcribe, fix the timings on the waveform, clear the reading-speed warnings, export SRT or WebVTT
  • Reach for it to turn hours-long lectures, films or podcast recordings into searchable, timed text without paying per minute
  • Reach for it when you have a spare NVIDIA GPU and want a self-hosted alternative to a per-minute transcription service

How Anysub compares

Anysub alongside other open-source audio, music & voice tools AI/TLDR tracks, ranked by GitHub stars.

ToolStarsWhat it does
Whisper★ 113kOpenAI's speech recognition model that transcribes and translates audio across many languages.
VoiceStudio★ 112kA local-first desktop voice studio that clones and designs voices, dubs video, and handles dictation, running 16 text-to-speech and 11 speech-recognition engines on your own hardware.
GPT-SoVITS★ 65.2kAn open-source WebUI that clones a voice from a short audio sample and turns text into speech, with zero-shot and few-shot fine-tuning.
Voicebox★ 60.7kLocal-first voice studio that clones a voice from a short sample, generates speech across seven TTS engines and 23 languages, handles system-wide dictation, and speaks for agents over MCP.
VibeVoice★ 57.2kMicrosoft's text-to-speech model for generating long, expressive multi-speaker audio like podcasts.
whisper.cpp★ 55.2kA dependency-free C/C++ port of Whisper built on ggml, running speech recognition on CPU, Metal, CUDA, Vulkan and NPUs from phones to servers.
Coqui TTS★ 46.2kA library of text-to-speech models including the multilingual XTTS voice-cloning model.
Anysub★ 3.1kA self-hosted subtitle studio (formerly Whishper) that transcribes audio and video on your own machine