MioSub turns a video link into translated, burned-in subtitles

The whole subtitle chain in one desktop window

The pitch is a single pipeline rather than a collection of tools. The project description calls it a one-stop fully automated subtitle generator that covers downloading, transcription, translation and burn-in with no human intervention, and the README makes the same point in a more relatable way, contrasting a workflow where you jump back and forth between transcription, translation, proofreading and timeline adjustment with a version where you paste a link, get a coffee, and come back to a finished file. The feature table spells out what is in that pipeline. Alignment is built in through a CTC aligner with millisecond precision, so there is no separate external tool to configure. Term extraction is automatic, and so is speaker labelling. The interface is Chinese and English with Japanese, and the project claims support for translation between more than a hundred languages. It also handles more than video, with podcasts, radio and audiobooks processed directly. The user is a person who publishes translated video and wants the tedious parts removed, not a developer building a speech pipeline.

The stack is a normal modern Electron app

The technology choices are visible in the badges and in the repository layout, and they are conventional. The interface is React 19, the shell is Electron 39, the language is TypeScript 5.8 and the styling is Tailwind CSS 4, with Vite as the bundler and electron-builder for packaging. The tree matches that stack, with src/ for the renderer, an electron/ directory for the main process, two Vite configuration files, vite.config.ts and vite.config.electron.ts, tailwind.config.js and postcss.config.js, a docs/ site built with fumadocs and Next, an openspec/ directory, a deploy/ directory, and.husky/ for git hooks alongside eslint.config.mjs and a.prettierrc. The package metadata shows the app is private, with the name miosub, the description Smart, Playful, and Powerful AI Subtitle Generator and the version 3.1.9, and the entry point of the built main process at dist-electron/main.cjs. Nothing here is exotic, which is a good sign for anyone who wants to read the code or patch the build.

The model access is the configuration decision that matters

The quick start has three steps. You download the installer for your platform from the releases page, you open the settings and fill in your Gemini API key along with a Gemini endpoint, and then you paste a video link or import a local file. The endpoint field takes a domain only, for example a URL of the form https://example.com, and the README also points to a full guide for running local Whisper. The key requirement is stated as a tip rather than buried: the API key needs to support the Gemini 2.5 and 3 Flash and Gemini 3 Pro models. The project also recommends a third party Gemini relay service, describing it as inexpensive and stable with broad model support, which is a reasonable convenience for users in regions where the direct endpoint is awkward, and it is a reminder that transcription and translation in this application run on someone else's server unless you point it at a local model. The environment template shows the full set of keys the application expects:

GEMINI_API_KEY=your_gemini_api_key_here
OPENAI_API_KEY=your_openai_api_key_here

VITE_AMPLITUDE_API_KEY=your_amplitude_api_key_here
VITE_MIXPANEL_TOKEN=your_mixpanel_token_here
VITE_SENTRY_DSN=your_sentry_dsn_here

The first two are the model credentials, and the VITE_ prefixed ones are product analytics and error reporting keys, which is worth noticing before you ship a build of your own.

Alignment, diarization and terminology are the accuracy story

The v3.0 highlights section is the part of the README that explains what changed and why. The headline feature is timeline alignment that works out of the box, with a built in CTC aligner giving millisecond precision and removing the need to configure a separate external tool. The second is a rebuilt editor and settings panel. The third is that the application is no longer limited to video, so pure audio files can be processed directly. The fourth is cleaner Chinese and Japanese rendering, with the NotoSans font bundled so that missing glyphs and tofu boxes stop appearing. Underneath, the accuracy claims are the automatic term extraction, the millisecond alignment and the automatic speaker labelling, and the demo table gives a sense of the range, with a 30 minute Japanese radio programme with speaker labels, a Japanese rail vlog heavy on specialist terminology, and a fast paced film commentary video with lots of character, group, producer and song names. Forced alignment is what makes the difference between a transcript and usable subtitles, and the credits also list whisper.cpp, which is the same idea implemented in C++ for local use.

The editor and the export formats

The application is not a batch script with no review step, and the screenshots section describes a working editor: live subtitle preview, automatic scrolling, speaker labels, floating playback, and search and replace. The import and export row of the feature table covers the formats that matter in practice, with SRT and ASS import, bilingual subtitle export, and one click burn-in of the finished file. ASS is the format that subtitle groups use for styling and positioning, so supporting it on import is a real convenience, and the topic tags mention substation alpha alongside srt subtitles, forced alignment, ffmpeg and diarization, which matches the credits where FFmpeg appears alongside yt-dlp for downloading and the various speech engines. The underlying Whisper model can be swapped for a local build, which is the escape hatch for people who do not want media leaving the machine.

Running the app from source

The local development section is short and assumes Node.js 18 or newer. The three commands are:

yarn install
yarn electron:dev
yarn electron:build

After packaging, the installers are found in the release directory. The package scripts show what sits behind those commands, and the detail that matters for a contributor is the preinstall and postinstall pair. The preinstall hook rewrites git URLs so that ssh://[email protected]/ is transparently replaced with an https://github.com/ URL, which is there so a clone works behind firewalls and in CI without key configuration. The postinstall hook runs node scripts/download-binaries.mjs, which fetches the external binaries the application depends on, and there are separate download:binaries and download:binaries:force variants of that step. The main process and renderer are built separately, with build:main using the dedicated vite.config.electron.ts configuration, and electron:build chains typecheck, the main process build, the renderer build and electron-builder. There are also i18n scripts for extracting and checking Chinese strings, which is how the translated interface is kept consistent.

Licence, renaming and release pace

The project is licensed under AGPL-3.0, and the package metadata records the same as AGPL-3.0-only with the licence file referenced from the package. That licence is the one to think about carefully, because the AGPL requires source availability for modified versions that are networked or distributed, and for a desktop application that means any build you hand to someone else is a build you are obliged to share the corresponding source of. It is a deliberate choice for a tool that is largely a wrapper around other people's work, Gemini and Whisper, and it is the opposite of what you would want in a component you intend to embed in a closed product. The project was renamed, since the README still carries the Gemini-Subtitle-Pro links and the release assets live under that name, and the header says formerly Gemini-Subtitle-Pro. The releases show quick patch cadence, with v3.1.7 on 2026-06-05, v3.1.8 and v3.1.9 both on 2026-07-18, with the most recent push to the repository coming after those releases. The repository is not archived, the documentation lives at miosub.app/docs with an English version in docs/README_en.md, and there is a hosted demo at demo.miosub.app for trying the workflow without installing anything.

Editorial conclusion

MioSub suits a subtitle creator or localisation team that produces translated video for multiple languages and wants the whole chain in one desktop application instead of a transcription service, a translation tool and a separate burn-in step. It is a poor fit for a team that needs to run this on a server at scale, since the repository marks the package private and the product is delivered as desktop installers, and a poor fit if you need results without a cloud model, since the default path depends on a Gemini API key. Verify first that your API key supports the Gemini 2.5 and 3 Flash and Gemini 3 Pro models the README requires, decide between a hosted endpoint and local Whisper before you start, and note that the AGPL-3.0 licence obliges you to publish source if you distribute a modified build.

Frequently asked questions

Official sources

1. corvo007/MioSub on GitHub
2. License: AGPL-3.0
3. Project website
4. README
5. Releases