WhisperSubTranslate

July 13, 2026 · View on GitHub

Turn any video into multilingual subtitles, locally. Drop in a video, generate an SRT with whisper.cpp, then translate it offline with the bundled Hy-MT2 model or with free/paid online engines.

This app creates new subtitles from your video's audio (speech to text). It does not extract embedded subtitle tracks or read on-screen text (no OCR).

Features

100% local speech to text. Your video never leaves your machine, no account, no upload.
Offline translation with the bundled Hy-MT2 model, or online engines (MyMemory, DeepL, OpenAI, Gemini) with your own keys.
Automatic model download. No Python, no manual setup.
Sync repair models (large-v2 Sync and Sync Lite) for videos where normal models drift out of sync.
Queue, live progress, and local-only job history.

Getting started

Users

Download the latest portable archive from Releases, extract it, and run WhisperSubTranslate.exe. Extraction runs fully offline on your PC. Translation is optional.

Developers

npm install
npm start

Node.js >= 20.19 or >= 22.12 (Electron 42 toolchain)
whisper.cpp is downloaded during npm install (CUDA build on Windows, ~700MB)
FFmpeg is included via npm; the selected GGML model downloads on first use

Linux

sudo apt install cmake build-essential git ffmpeg   # Ubuntu/Debian
npm install   # whisper.cpp is built from source
npm start

For CUDA acceleration, install the NVIDIA CUDA Toolkit before npm install. Manual whisper.cpp build steps are in CONTRIBUTING.md.

Build (Windows)

npm run build-win   # artifacts are emitted to dist2/

Translation engines

Translate subtitles fully offline with the bundled Tencent Hy-MT2 model, or route to free/paid online engines using your own API keys.

Engine	Offline	API key	Cost	Notes
Hy-MT2 1.8B (local, default)	Yes	No	Free	~1.13GB, VRAM 2GB / RAM 4GB, on-device
Hy-MT2 7B (local)	Yes	No	Free	~6.16GB, VRAM 8GB / RAM 12GB, larger model
MyMemory	No	No	Free	~50K chars/day per IP
DeepL	No	Yes	Free 500K/month	Deterministic output
OpenAI GPT-5.4 mini	No	Yes	Paid	Context-aware
OpenAI GPT-5.4 nano	No	Yes	Paid	Cheaper tier
Gemini 3 Flash	No	Yes	Free / low-cost	Recommended low-cost route (get key)

The local Hy-MT2 engine is the only option that needs no API key, no network, and no per-use cost, so your dialogue never leaves your machine.

Translation quality (offline engine)

WhisperSubTranslate ships Tencent's Hy-MT2 models (1.8B default, 7B optional). Tencent's official evaluation shows the Hy-MT2 family competing with leading commercial translation APIs, and ahead of several of them on some benchmarks.

Hy-MT2 translation benchmark, official Tencent figures, bundled in WhisperSubTranslate

Source: official benchmarks from Tencent: Hy-MT2 repository, technical report, models on HuggingFace. The chart is redrawn from Tencent's official Figure 1, with bundled-model (1.8B/7B) numbers checked against the paper tables. These figures measure the underlying model on standard machine translation benchmarks (WildMTBench, WMT25, FLORES-200, etc.), not a WhisperSubTranslate-specific benchmark.

For long videos (1hr+), MyMemory's daily limit can cause slowdowns. Use Gemini, DeepL, or a configured GPT model instead.

Speech recognition models

Models download on demand into _models/. CUDA is used when available, otherwise CPU runs by default. Pick a size that fits your GPU.

Model	Size	VRAM	Speed	Notes
tiny	~75MB	~1GB	Fastest	Basic
base	~142MB	~1GB	Fast	Good
small	~466MB	~1GB	Medium	Better
medium	~1.5GB	~2GB	Medium	Great
large-v3	~3GB	~4GB	Slow	Best transcription
large-v3-turbo (default)	~809MB	~2GB	Fast	Best all-round
large-v2 Sync	~4.4GB	~4.5GB	Slow	Separate engine; fixes subtitle sync
large-v2 Sync Lite	shared	~3GB	Slow	Same file as Sync, int8, lower VRAM

Sync and Sync Lite use a separate Faster-Whisper engine (auto-downloaded once, ~4.4GB) and share the same model file, so one download covers both. Use them only when normal models drift out of sync; they are most accurate on non-English video (Japanese, Korean, Chinese). English is usually fine with large-v3-turbo.

VRAM figures for whisper.cpp models are with GGML optimization, much lower than PyTorch Whisper (~10GB for large). Sync figures are from the Faster-Whisper benchmark.

Language support

UI: Korean, English, Japanese, Chinese, Polish
Translation targets (15): ko, en, ja, zh, es, fr, de, it, pt, ru, hu, ar, pl, tr, fa
Audio recognition: 100+ languages via whisper.cpp

Data storage

Everything stays local under your user data folder. Nothing is uploaded.

Data	Location
Settings & API keys	`%APPDATA%\whispersubtranslate\translation-config-safe.json`
Job history	`%APPDATA%\whispersubtranslate\history.json` (up to 200 entries)
Error logs	`%APPDATA%\whispersubtranslate\logs\errors.log`
Models	`_models/` (in app folder)

API keys are stored locally with OS-level safe storage, and the config is never committed or bundled. Job history is optional (toggle in Settings) and capped at 200 entries.