feat: add local multilingual voice input

Author: JeffGuo7

.github/workflows/ci.yml

@@ -31,3 +31,22 @@ jobs:
 
       - run: npm run test:server
         name: Run pi-server tests
+
+  windows-native:
+    runs-on: windows-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: dtolnay/rust-toolchain@stable
+
+      - name: Configure LLVM
+        shell: pwsh
+        run: |
+          "LIBCLANG_PATH=C:\Program Files\LLVM\bin" >> $env:GITHUB_ENV
+
+      - run: cargo test --manifest-path stt/Cargo.toml --lib
+        name: Test local speech engine
+
+      - run: cargo check --manifest-path src-tauri/Cargo.toml
+        name: Check Tauri native layer

.gitignore

@@ -59,6 +59,8 @@ tmp_*
 
 # Test coverage output
 coverage/
+/stt/target/
+/stt/Cargo.lock
 
 # Windows temp file
 nul

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,11 @@
+# Third-party notices
+
+## OpenWorker local speech-to-text design
+
+The structure and portions of the local speech-to-text implementation in `stt/` were adapted from
+OpenWorker's `ocw-stt` crate.
+
+Copyright (c) 2024 Andrew Ng
+
+Licensed under the MIT License. The full license text is available in the OpenWorker source
+distribution and at <https://opensource.org/license/mit>.

docs/voice-input.md

@@ -0,0 +1,46 @@
+# Local voice input
+
+Snotra's first voice feature intentionally matches OpenWorker's current scope: local dictation
+only. It does not implement text-to-speech, a voice assistant loop, partial streaming transcripts,
+or remote audio processing.
+
+## Architecture
+
+- `stt/` is a Tauri-independent Rust crate that owns microphone capture, model provisioning, and
+  final Whisper transcription.
+- `src-tauri/src/dictation.rs` exposes the engine as Tauri commands and reports compatibility.
+- `src/utils/voiceInput.ts` is the typed frontend boundary.
+- Settings own model download, checksum verification, microphone testing, and deletion.
+- The chat input owns the recording UX. A transcript is appended to the editable draft and is
+  never sent automatically.
+
+Audio samples remain in memory for the active recording and are discarded after transcription or
+cancel. The only persistent files are the downloaded model and small verification/readiness
+markers in the app data `models` directory.
+
+## Model
+
+Snotra uses multilingual Whisper Base (`ggml-base.bin`) so Chinese and English can be detected
+automatically. The model is downloaded on explicit user action and is not bundled into `.exe`,
+`.msi`, or `.dmg` installers.
+
+- Download size: `147951465` bytes
+- SHA-256: `60ed5bc3dd14eea856493d334349b405782ddcaf0028d4b5df4088345fba2efe`
+- Source: `ggerganov/whisper.cpp` on Hugging Face
+
+Downloads are written to a `.part` file, can be cancelled, and only replace the active model after
+both the expected size and SHA-256 pass.
+
+## Native build requirements
+
+`whisper-rs` compiles whisper.cpp during the Rust build. Build machines need:
+
+- Rust with the MSVC target on Windows
+- Visual Studio C++ Build Tools and a Windows SDK
+- CMake
+- LLVM/libclang (set `LIBCLANG_PATH` if it is not discoverable)
+
+These are build-time requirements only. End users do not need Node.js, Git, Python, LLVM, CMake, or
+Visual Studio to use voice input in an installed Snotra build.
+
+macOS builds also merge `src-tauri/Info.plist`, which contains the microphone usage description.