SmartSub
buxuku/SmartSubReviewed by AI
Checked by RepoPicks Radar on
What is SmartSub?
SmartSub (妙幕) is a desktop app for subtitles. You drop in video or audio files, or paste a link from YouTube or Bilibili, and it writes down what is said using speech models that run on your machine, translates the subtitles, lets you correct them line by line against the video, can read the translation out loud in a synthetic or cloned voice, and finally burns or attaches the subtitles to the video. There are installers for Windows, Mac and Linux, and the first run asks you to download a speech model. Cloud services are optional extras.
SmartSub is free and open source (MIT) and can replace Descript, ElevenLabs, and Otter.ai.
How easy is it to install SmartSub?
B · installer plus one simple step · Windows
- Windows B
- macOS B
- Linux B
All three systems get B: there is a ready installer for each (a signed .exe for Windows, .dmg for Intel and Apple-chip Macs, .deb or AppImage for Linux, plus a Homebrew cask), and the one extra step is the first-run wizard asking you to download a speech model, with GPU acceleration packs also fetched inside the app. Windows is the smoothest. On a Mac the FAQ warns the app may be reported as damaged and tells you to run "sudo xattr -dr com.apple.quarantine /Applications/SmartSub.app" once in the terminal.
SmartSub on GitHub
- Stars
- 6,000on Nov 10, 2026
- Downloads
- 479,4119,778 this week
- Growth
- +57stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Nothing extra
- To build from source
- Node.js, NVIDIA graphics card, FFmpeg
- Disk space
- 108 MB–324 MB · from the installer
- Accounts or keys
- AI models for translating, polishing and the built-in assistant: DeepSeek, Qwen, Gemini, SiliconFlow or any OpenAI-compatible endpoint (Ollama on your own computer works too), Cloud speech-to-text, nine providers including OpenAI, Deepgram, ElevenLabs Scribe, Gladia, Tencent, Alibaba, iFlytek, Volcengine and Xiaomi MiMo, Cloud dubbing voices: Azure Speech, ElevenLabs, Volcengine Doubao, OpenAI-compatible endpoints
- Cost
- Free · The app is free and MIT licensed, and the README describes a route that costs nothing: local speech models, the built-in free translation sources, local voices and the built-in ffmpeg. The paid cloud options (OpenAI, ElevenLabs, Azure, Tencent, Alibaba and others) are billed by those companies if you turn them on.
- License
- MIT (Permissive)
- Can replace
- Descript, ElevenLabs, Otter.ai
- Last change on GitHub
- Nov 9, 2026
SmartSub risk signals (8)
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): diff: GHSA-73rr-hh4g-fpgx
automatic check
- Attentiondownloads videos and unlocks paid content with cookies
The app has a built-in downloader with "yt-dlp(YouTube 及 1800+ 站点)与 lux(B 站、抖音、小红书等国内平台)双引擎", and it can import your browser cookies to "解锁登录后可见的高清画质与会员内容" — unlock the higher quality and members-only content you can see once logged in. Saving videos this way is against the terms of YouTube, Bilibili and similar sites, and in some countries it is restricted beyond that. Use with your own material or material you are allowed to copy.
AI review
- Attentioncopies a real persons voice
The dubbing part offers "声音克隆:本地 ZipVoice 零样本克隆(一段参考音频即建即用)" plus ElevenLabs instant cloning and Volcengine voice replication 2.0: one short recording is enough to make a video speak in that voice. Nothing checks whether the voice's owner agreed, and the output is not marked as generated.
AI review
- Infofree voice service is a reverse engineered endpoint
In the dubbing engines table, Edge TTS is described as "免费、无需 Key;属逆向接口试用档,不承诺可用性,断供时建议切换其它引擎" — free with no key because it speaks to a Microsoft interface that was reverse engineered, with no promise it keeps working.
AI review
- Infothe page earns from the services it suggests
The sponsor block at the top of README.md advertises an API reseller with a referral link (aiyoyoo.com/register?aff=YWL9FL2XFPKM, "倍率低至0.05x", "新用户注册赠送 $20"), the translation list links DeerAPI with ?aff=QvHM, and the sponsors section uses a DigitalOcean referral link. So some recommendations inside the project page pay the author.
AI review
- Infomac build is not notarized
Only the Windows installer is signed: the code signing section says the scope is "Windows 安装包(SmartSub_Windows_<版本>_x64.exe)" through SignPath, and "签名从证书签发后的首个版本开始生效,此前发布的版本未签名". The Mac FAQ needs the quarantine flag removed by hand, which is what an unsigned, un-notarized app looks like.
AI review
- Infobackground service that outside assistants can drive
The automation section describes a "无窗口后台守护(Headless Daemon)" that starts by itself the first time an AI tool or the CLI calls it, and says the README's 111 MCP tools let Cursor, Claude Code or Codex run the whole pipeline. An assistant you connect can then read and write your media files and the app's settings, including the keys stored there.
AI review
- Infomodels come from several download sources
Models are fetched inside the app from Hugging Face, the hf-mirror.com mirror, ModelScope and GitHub ("支持 ModelScope / GitHub 等多源"), and the download engines yt-dlp and lux are installed from inside the app as well. The 108 MB installer is only the start; the model files are much bigger.
AI review
Evidence is shown in its original language.
What Radar could not check about SmartSub
- How accurate the subtitles are for a given language, and how long one hour of video takes on your machine
- Real disk space after the speech models, voice models and GPU acceleration packs are downloaded
- Whether the GPU acceleration really works on your graphics card, since the FAQ already answers crashes after turning it on
- The privacy claim that nothing is uploaded: only the parts you point at a cloud service can be checked from outside
- What the cloud providers charge for long videos
- Whether the cloned voice is good enough to pass for the original