Local & Online Speech-to-Text for Windows
简体中文 | English
Press a hotkey, speak, release — text appears at your cursor.
Download 轻语.Whisper_x.x.x_x64-setup.exe from the Releases page. Run the installer — everything is bundled, no Python or build tools needed. ASR models will be downloaded on first use (~1–1.5 GB).
Note
GPU (optional): NVIDIA GPU with up-to-date driver for CUDA acceleration. No GPU → automatic CPU fallback.
See Quick Start below.
|
One-key dictation Three ASR engines AI polish Adaptive learning Subtitle overlay |
Voice assistant Edit selected text Real-time translation & more |
| Feature | Light-Whisper | Typeless |
|---|---|---|
| Pricing | Free & open-source | Free tier (4k words/week); $12–30/mo |
| Privacy | Fully offline (local engines) | Cloud-based, zero-data-retention |
| Open source | ✅ | ❌ |
| Platform | Windows | Windows, Mac, iOS, Android, Web |
| ASR engines | 3 switchable (local + online) | Cloud proprietary |
| Languages | 5–99+ (engine dependent) | 100+ |
| AI polish | Multi-backend LLM, bring your own key | Built-in |
| Screen-aware assistant + web search | ✅ | ❌ |
| Subtitle overlay | ✅ | ❌ |
| SenseVoice (default) | Faster Whisper | GLM-ASR (online) | |
|---|---|---|---|
| Chinese CER | 2.96 % (AISHELL-1) | 5.14 % | 7.17 % |
| English WER | 3.15 % (LibriSpeech) | 1.82 % | — |
| Languages | 5 (zh/en/ja/ko/yue) | 99+ | Chinese + dialects |
| Punctuation | Built-in ITN | initial_prompt guided | Built-in |
| Hot words | ✅ | ✅ | ✅ (max 100) |
| Model size | ~938 MB | ~1.5 GB | Cloud (no download) |
| Requires | GPU/CPU (bundled) | GPU/CPU (bundled) | API key only |
| Cost | Free (local) | Free (local) | ¥0.06/min |
Note
SenseVoice/Whisper CER source: FunAudioLLM paper, Table 6. GLM-ASR CER source: Zhipu AI.
Important
Windows 10/11 (x64) only. Disk: ~10 GB free. These requirements are only needed for building from source — the installer bundles everything.
| Tool | Version | Purpose |
|---|---|---|
| Visual Studio Build Tools | 2019+ | MSVC C++ toolchain |
| Rust | >= 1.75 | Backend |
| Node.js | >= 18 | Frontend build |
| pnpm | >= 8 | Frontend packages |
| uv | >= 0.4 | Python env (auto-installs Python 3.11) |
GPU (optional): NVIDIA GPU with up-to-date driver. No need to install CUDA Toolkit — PyTorch bundles CUDA 12.8.
Tip
GLM-ASR users: If you only use the online GLM-ASR engine, you only need Rust, Node.js, and pnpm — no Python, uv, or GPU required. Just build and add your API key in Settings.
Step-by-step tool installation
# 1. Visual Studio Build Tools — download installer, check "Desktop development with C++"
# 2. Rust
winget install Rustlang.Rustup
# 3. Node.js + pnpm
winget install OpenJS.NodeJS.LTS
npm install -g pnpm
# 4. uv
winget install astral-sh.uvVerify:
rustc --version # >= 1.75
node --version # >= 18
pnpm --version # >= 8
uv --version # >= 0.4[!TIP] Python is managed by
uv— no manual install required.
git clone https://github.com/sypsyp97/light-whisper.git
cd light-whisper
pnpm install # Frontend deps
uv sync # Python deps (downloads Python 3.11, PyTorch CUDA, etc. ~5-15 min)# SenseVoice (default, ~938 MB)
uv run python -c "from huggingface_hub import snapshot_download; snapshot_download('FunAudioLLM/SenseVoiceSmall'); snapshot_download('funasr/fsmn-vad')"
# Faster Whisper (~1.5 GB)
uv run python -c "from huggingface_hub import snapshot_download; snapshot_download('deepdml/faster-whisper-large-v3-turbo-ct2')"Models are cached in ~/.cache/huggingface/hub/.
Tip
China mainland: set $env:HF_ENDPOINT = "https://hf-mirror.com" before downloading.
pnpm tauri build # First build ~5-15 min (compiles Rust deps)The installer is in src-tauri/target/release/bundle/nsis/, or run src-tauri/target/release/light-whisper.exe directly.
┌──────────────┐ ┌──────────────┐ ┌─────────────────┐
│ React UI │ Tauri IPC │ Rust Core │ stdin/stdout │ Python ASR │
│ TypeScript │◄──invoke/emit─►│ (Tauri 2) │◄────JSON──────►│ SenseVoice / │
└──────────────┘ └──────┬───────┘ │ Faster Whisper │
│ └─────────────────┘
├─── HTTP ──► GLM-ASR API (online ASR)
├─── HTTP ──► LLM API (AI polish / assistant / translation)
├─── HTTP ──► Web Search (Exa / Tavily) → assistant context
├─── Screen Capture ──► full-screen screenshots → assistant context
└─── User Profile ──► hot words + blacklist → ASR + LLM prompt
Key paths
| Layer | Paths |
|---|---|
| Frontend | src/pages/, src/components/, src/hooks/, src/contexts/, src/lib/, src/i18n/, src/styles/ |
| Rust commands | src-tauri/src/commands/ — audio, assistant, clipboard, funasr, hotkey, ai_polish, profile, updater, window |
| Rust services | src-tauri/src/services/ — funasr_service, glm_asr_service, audio_service, assistant_service, ai_polish_service, llm_client, llm_provider, profile_service, screen_capture_service, web_search_service, download_service |
| State | src-tauri/src/state/ — app_state, user_profile |
| Python ASR | src-tauri/resources/ — funasr_server.py, whisper_server.py, server_common.py |
pnpm tauri dev # Dev mode with hot-reload
pnpm tauri build # Production build + installer
pnpm build # Frontend only
uv sync # Sync Python deps
cd src-tauri && cargo check # Rust type checkPyTorch or model downloads are slow
- PyTorch CUDA (~2.5 GB) downloads from
download.pytorch.org— use a stable connection or VPN.uv syncsupports resume. - Other Python packages can use a Tsinghua mirror:
$env:UV_INDEX_URL = "https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple" - HuggingFace models:
$env:HF_ENDPOINT = "https://hf-mirror.com"
GPU not detected
Verify: .venv\Scripts\python.exe -c "import torch; print(torch.cuda.is_available())" should print True.
Requires an up-to-date NVIDIA driver. For CUDA 12.x minor-version compatibility, Windows driver should be >= 528.33. The app falls back to CPU automatically if no GPU is found.
Characters turn into periods when typing into apps
This happens when a Chinese IME intercepts SendInput Unicode events. Fix: switch to Clipboard paste mode in Settings, or toggle your IME to English mode.
Chinese text appears garbled
Enable Windows UTF-8 system locale support to resolve encoding issues:
Control Panel → Region (or Clock and Region) → Administrative → Change system locale... → check Beta: Use Unicode UTF-8 for worldwide language support → OK → restart Windows.
Hotkey not working
Default dictation hotkey is F2. If occupied by another program, change it in Settings (e.g. Ctrl+Win+R).
Log locations
%APPDATA%\com.light-whisper.app\logs\ — funasr_server.log / whisper_server.log
- FunASR & SenseVoiceSmall — Alibaba DAMO Academy
- faster-whisper & large-v3-turbo-ct2
- GLM-ASR — Zhipu AI
- Tauri / React
Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)