WinLocalASR

Offline voice dictation for Windows 10+ (x64). Press a global hotkey, speak, and your words are copied to the clipboard — no cloud, no internet after first-time setup.

WinLocalASR 是一款 Windows 10+（x64）离线语音听写工具。按下全局热键、说话，转写文本自动进剪贴板；首次配置后完全离线运行，不依赖任何云服务。

Features

• Lives in the system tray — no window, no taskbar clutter（常驻系统托盘，无主窗口）
• Global hotkey (default Ctrl+Shift+Space) toggles recording; Esc cancels mid-recording（全局热键切换录音，录音中 Esc 取消）
• 24 kHz mono capture with a live level meter（24 kHz 单声道采集 + 实时电平表）
• Local transcription via Qwen3-ASR-1.7B (GGUF, Q8_0) on llama.cpp — CPU with AVX2 is enough, no GPU required（本地 llama.cpp 推理，AVX2 CPU 即可运行，无需 GPU）
• Transcript goes to the clipboard — paste with Ctrl+V anywhere（转写结果进剪贴板，随处粘贴）
• HUD overlay feedback: recording timer + level bars, processing spinner, copied / cancelled / error states（悬浮 HUD 实时反馈）
• Configurable: hotkey, input device, hotwords/context prompt, recording limit (10–120 s), autostart（热键/设备/热词/录音上限/开机自启均可配置）
• Bilingual UI: English + 简体中文, following system locale（界面双语，跟随系统语言）
• First-run setup downloads the engine + model once (~2.5 GB) with automatic China-friendly mirrors and interrupted-download resume（首装自动下载，国内网络自动走镜像，支持断点续传）

Status: v0.1.0 prerelease. The full pipeline (build, tests, end-to-end run, real-inference smoke, installer regression) is verified on every push by GitHub Actions. Validation on physical hardware is still in progress — see docs/test.md §Real-Machine.

Requirements

• Windows 10 or later, x64
• CPU with AVX2
• ~2.6 GB free disk space (engine ~18 MB + model ~2.5 GB)

Install

Option A — installer (recommended)

1. Download WinLocalASR-Setup-vX.Y.Z-x64.exe from Releases.
2. Run it. The binary is not code-signed, so Windows SmartScreen may show "Windows protected your PC". Click More info → Run anyway.（未购买代码签名证书；SmartScreen 提示「Windows 已保护你的电脑」时，点击「更多信息 → 仍要运行」。）
3. Launch WinLocalASR from the Start menu. First launch opens a guided setup that downloads the engine + speech model (~2.5 GB; a China-friendly mirror is used automatically when the direct download fails).（首次启动引导下载模型，国内网络自动走镜像，支持断点续传。）
4. Grant microphone permission when Windows asks.（按系统提示授权麦克风。）

Option B — portable zip

1. Download WinLocalASR-vX.Y.Z-portable-win-x64.zip from Releases and extract it anywhere.
2. Run WinLocalASR.bat (or WinLocalASR.App.exe directly). SmartScreen guidance applies the same as above.
3. The same first-run setup runs on first launch.

The app is fully offline after setup completes — you can disconnect the network.

Usage

1. WinLocalASR sits in the system tray. Left-click the tray icon to open the menu:
   • Setup… — run/repair first-time setup (bold while not configured)
   • Settings… — hotkey, input device, hotwords, recording limit, autostart
   • Copy last transcript — re-copy the previous transcript
   • Restart engine — restart the local inference server
   • Exit（托盘菜单：Setup / 设置 / 复制上次转写 / 重启引擎 / 退出）
2. Press Ctrl+Shift+Space (or your custom hotkey). A red HUD appears in the bottom-right corner with a timer and level bars — speak now.（按下热键，右下角出现红色 HUD 与电平表，开始说话。）
3. Press the hotkey again to stop. The HUD turns amber while transcribing, then green ✓ — the text is on the clipboard, paste with Ctrl+V.（再按一次停止；转写完成显示绿色 ✓，Ctrl+V 粘贴。）
4. Press Esc while recording to cancel and discard. Recording also auto-stops at the configured limit (default 120 s).（录音中按 Esc 取消；达到上限自动停止。）
5. Tip: put names or jargon into Settings → hotwords (context prompt) to bias recognition.（把专有名词填进「热词」可提高识别准确率。）

Note: while recording, Esc is captured system-wide (pressing it in any application cancels the recording). Outside recording, Esc is never touched. This mirrors the macOS original — see docs/prd.md "Known tradeoffs".

Troubleshooting

• Transcription fails with a microphone error / level meter stays flat（麦克风被拒/电平表无反应）: Windows may be blocking microphone access for desktop apps. Open Settings → Privacy & security → Microphone and enable both "Let apps access your microphone" and "Let desktop apps access your microphone", then restart WinLocalASR. Check that the correct input device is selected in Settings.
• The hotkey does nothing / a balloon says the hotkey is occupied（热键无效/托盘气泡提示冲突）: Another application has claimed the combination. WinLocalASR keeps the previously working hotkey and shows the conflict as a balloon plus a red hint in Settings. Open Settings → hotkey, press a different combination (any Ctrl/Alt/Shift/Win + key), Apply.（其他程序占用了组合键；在设置里换一个组合即可，冲突会在设置页标红提示。）
• Transcription is slow on a low-end CPU（低配 CPU 转写慢）: All transcription runs locally on your CPU; on low-end machines the wait after stop will be longer than on the reference machine. Close CPU-heavy applications (browsers with many tabs, games, video encoders) while dictating. The model ships only in the Q8_0 quantization (the official GGUF release provides no smaller quant), so there is no "lighter" official variant to switch to; expect proportionally longer waits for longer recordings.（转写完全在本地 CPU 进行；关闭高负载程序可明显改善。官方仅提供 Q8_0 量化，没有更小的官方版本可换。）
• First-run setup download fails or is very slow（首装下载失败/极慢）: The setup automatically falls back to China-friendly mirrors (hf-mirror.com for the model, ghproxy.cn for the engine) and resumes interrupted downloads — just run Setup again from the tray menu; finished pieces are not re-downloaded.（下载中断后重跑 Setup 即可续传，已完成的文件不会重复下载。）
• Uninstall keeps ~2.5 GB on disk（卸载后仍有约 2.5 GB 占用）: By design the uninstaller leaves the downloaded model directory (%LOCALAPPDATA%\WinLocalASR) in place so reinstalls don't re-download; delete that folder manually if you want the space back.（卸载保留模型目录以便重装免下载；需要空间可手动删除 %LOCALAPPDATA%\WinLocalASR。）

FAQ

• Does it send my voice anywhere? No. After first-run setup, the network allowlist is localhost-only; audio is processed by a local llama.cpp process. You can verify by dictating with networking disabled.（完全离线；可断网验证。）
• Where are files stored? Engine + models: %LOCALAPPDATA%\WinLocalASR. Settings: %APPDATA%\WinLocalASR\settings.json.（引擎与模型在 %LOCALAPPDATA%\WinLocalASR，设置在 %APPDATA%\WinLocalASR。）
• Why does recording stop at 2 minutes? The recording limit is capped at 120 s as a guardrail against a known upstream issue with >2 min audio; 10–120 s is configurable in Settings.（录音上限 120 秒，可在 10–120 秒内调整。）
• Is there a GPU build / streaming mode? Not currently. CPU-only inference, batch transcription per utterance — no streaming, no VAD, by design.（暂无 GPU 构建；无流式/自动分段，属设计取舍。）
• Which languages? The underlying Qwen3-ASR model is multilingual (English and Chinese are the tested configurations); recognition quality for other languages follows the model itself.（模型多语种；英文与中文为已验证配置。）

Development

dotnet build src/WinLocalASR.sln
dotnet test src/WinLocalASR.sln

The App target is net10.0-windows WinForms and builds cross-platform via EnableWindowsTargeting; Windows-only tests run on windows-latest CI. CI also runs the ControlServer-driven e2e (tests/e2e.ps1), a real-inference smoke job on tags, and installer regression — see AGENTS.md and docs/test.md.

Credits

• grapeot/mac-native-local-asr — the macOS original this project ports to Windows (MIT). Its design shaped the hotkey flow, setup experience, HUD behavior, and documentation structure. See NOTICE.
• llama.cpp — local inference backend serving Qwen3-ASR.
• Qwen3-ASR (Alibaba Qwen Team) — the speech recognition model, ggml-org/Qwen3-ASR-1.7B-GGUF quantized release.

License

MIT — see LICENSE.