Blue-B/WhisperSubTranslate

A free, local desktop app to extract subtitles (SRT) from video and translate them into any language — unlimited use, no signup, no cloud.

WhisperSubTranslate – Quick Overview

What it does – Takes a video file, runs speech‑to‑text locally with whisper.cpp, then translates the generated subtitles. All heavy work can stay on your computer; you only need an internet connection the first time a model is downloaded.

Key points

Local transcription – No upload, no account. Uses Whisper models (tiny‑base‑large‑v3‑turbo, plus special Sync models for better timing on non‑English content).
Offline translation – Optional Tencent Hy‑MT2 models (1.8 B ≈ 1 GB, 7 B ≈ 6 GB). After download, translation runs completely offline, no API key needed.
Online translation – Supports MyMemory (free), DeepL, OpenAI, Gemini, Claude and custom OpenAI‑compatible endpoints if you prefer a cloud service.
Desktop app – Built with Electron + Node.js, packaged as a portable Windows ZIP (≈ 2 GB). Linux can be built from source.
Convenient UI – Queue, progress bar, job history, and a “portable‑data” mode that keeps all models and settings on a USB stick.

Why it matters – It bundles the Whisper speech‑recognition engine and a state‑of‑the‑art translation model into a single, easy‑to‑run program, letting anyone create multilingual subtitles without sending their video or text to a third‑party service.

How to get started (User view)

Download the Windows ZIP from the Releases page (≈ 2.15 GB). No Python or CUDA toolkit is required for the pre‑built package.
Extract the ZIP to a folder and run WhisperSubTranslate.exe.
Add a short video, pick the spoken language (or Automatic), leave Translation → No translation, and click Start Extraction.
The app will download the default Whisper model (large‑v3‑turbo, ~1.6 GB) the first time you run it. After transcription finishes, an .srt file appears next to the video.
To translate, choose Hy‑MT2 (local) or any online engine, select a target language, and run the job again (or load an existing SRT). The Hy‑MT2 model (1.8 B, ~1 GB) downloads on first use.

Disk & bandwidth – Expect about 3.8 GB of downloads for the app + default speech model, plus ~1 GB more if you add the 1.8 B translation model.

Technical highlights (Developer view)

Core stack – Electron front‑end, Node.js (≥ 22.12), whisper.cpp compiled with CUDA/Vulkan support, FFmpeg bundled via npm.
Model handling – Models are stored under %APPDATA%\whispersubtranslate (or portable-data/ for a self‑contained install). Integrity is verified with SHA‑256 hashes.
GPU support – Whisper can run on CUDA or Vulkan‑compatible GPUs; translation back‑ends also auto‑select CUDA/Vulkan when available, otherwise fall back to CPU.
Build – npm ci installs dependencies and builds whisper.cpp. Windows build: npm run build-win. Linux build requires cmake, build-essential, git, ffmpeg and optional CUDA toolkit.
Extensibility – UI strings live in locales/*.json and are translatable via Weblate. Adding new translation targets or UI languages follows the guide in docs/TRANSLATION.md.

Who might use this?

Content creators who need subtitles in multiple languages but cannot or do not want to upload raw video to cloud services.
Educators preparing multilingual lecture videos.
Privacy‑conscious users who want all audio/text processing to stay on‑device.

License & credits

License: GPL‑3.0 for the application code.
Core components: whisper.cpp (Georgi Gerganov), Tencent Hy‑MT2, FFmpeg, Faster‑Whisper, Silero VAD, and various npm packages.
Funding: GitHub Sponsors, Buy Me A Coffee, PayPal.