WhisTxtR (Whisper Tray) – Build & Run

Resident Windows tray app that records on hold-to-talk (Ctrl+Shift), runs Whisper locally (CPU or CUDA), and injects the transcript into the active app. Overlay shows recording/processing and live partial text.

Prerequisites (Windows 10/11)

• Rust (stable, MSVC toolchain).
• Visual Studio 2022 Build Tools (MSVC, Windows 10 SDK).
• CMake (tested path: G:\Unity Projects\Cmake\bin\cmake.exe).
• CUDA 12.4 for GPU builds (install toolkit; tested at C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4).
• libclang for bindgen (tested at J:\llvm\bin with LIBCLANG_PATH set there).

Quick env setup (PowerShell example)

# Adjust paths to your machine
$env:CMAKE           = 'G:\Unity Projects\Cmake\bin\cmake.exe'
$env:CMAKE_GENERATOR = 'Visual Studio 17 2022'
$env:LIBCLANG_PATH   = 'J:\llvm\bin'
$env:CUDA_PATH             = 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4'
$env:CUDA_PATH_V12_4       = $env:CUDA_PATH
$env:CUDA_TOOLKIT_ROOT_DIR = $env:CUDA_PATH
$env:CUDACXX               = "$env:CUDA_PATH\bin\nvcc.exe"
$env:CudaToolkitDir        = $env:CUDA_PATH
$env:PATH = "$env:CUDA_PATH\bin;$env:CUDA_PATH\lib\x64;$env:PATH"

Verify: where nvcc, cmake --version, where clang (or where libclang.dll).

Build

• CPU: cargo build --release
• CUDA: cargo build --release --features cuda

Outputs land in rust/target/release/stt-whis-tray.exe.

Run

From repo root:

cd J:\whistxt\rust
.\target\release\stt-whis-tray.exe

Config lives at %APPDATA%\WhisTray\config.json (created on first run).

Models (not in repo)

• Default expected path: J:/whistxt/models/ggml-small-q5_1.bin (set in config).
• Models are not committed; place your .bin/.gguf under models/ and update config if needed.
• Provide/download your own and verify checksum; add a small download script if distributing externally.

Hotkey & behavior

• Hold Ctrl+Shift to record; release to transcribe and inject.
• Overlay: topmost bar with red blink; shows “Listening…”, live partial text during processing, then final text.
• Injection: defaults to clipboard+Ctrl+V. Config is in %APPDATA%\WhisTray\config.json (inject_mode, append_newline, model_path, use_cuda, etc.).
• Tray menu: right-click tray icon → Quit, toggle “Start with Windows” (startup link).

Troubleshooting

• Missing CUDA DLLs: ensure CUDA_PATH/bin and CUDA_PATH/lib/x64 are on PATH; correct CUDA_TOOLKIT_ROOT_DIR/CUDACXX.
• Bindgen/libclang errors: set LIBCLANG_PATH to your LLVM bin folder.
• Model not found: confirm the path in %APPDATA%\WhisTray\config.json exists.
• Partial overlay not updating: ensure the app is running from the same directory as the model and CUDA DLLs are visible on PATH; overlay draws via Win32/GDI (no GPU UI dependency). Use the log output in the console to confirm state changes.

What’s not included

• Model binaries (large): ignored by .gitignore.
• Any network calls: audio and text stay on-device.

License: MIT (matches Cargo.toml).