[WIP] Add Windows support for meeting audio transcription

.github/workflows/clips-desktop-build-check.yml

@@ -58,6 +58,20 @@ jobs:
       - name: Install dependencies
         run: pnpm install --frozen-lockfile
 
+      # whisper-rs compiles whisper.cpp (CMake + C/C++) and runs bindgen, which
+      # needs libclang. windows-latest ships LLVM and CMake preinstalled; point
+      # bindgen at libclang and fall back to a Chocolatey install if missing.
+      - name: Set up libclang for whisper-rs bindgen (Windows)
+        if: runner.os == 'Windows'
+        shell: bash
+        run: |
+          if [ -f "/c/Program Files/LLVM/bin/libclang.dll" ]; then
+            echo "LIBCLANG_PATH=C:\\Program Files\\LLVM\\bin" >> "$GITHUB_ENV"
+          else
+            choco install llvm --no-progress -y
+            echo "LIBCLANG_PATH=C:\\Program Files\\LLVM\\bin" >> "$GITHUB_ENV"
+          fi
+
       - name: Build frontend
         working-directory: ${{ env.TAURI_APP_PATH }}
         run: pnpm run build

.github/workflows/clips-desktop-release.yml

@@ -123,6 +123,21 @@ jobs:
           V=$(node -p "require('./${{ env.TAURI_APP_PATH }}/package.json').version")
           echo "version=$V" >> "$GITHUB_OUTPUT"
 
+      # whisper-rs compiles whisper.cpp (CMake + C/C++) and runs bindgen, which
+      # needs libclang. windows-latest ships LLVM and CMake preinstalled; point
+      # bindgen at libclang and fall back to a Chocolatey install if the image
+      # ever drops it. macOS already has clang via Xcode.
+      - name: Set up libclang for whisper-rs bindgen (Windows)
+        if: runner.os == 'Windows'
+        shell: bash
+        run: |
+          if [ -f "/c/Program Files/LLVM/bin/libclang.dll" ]; then
+            echo "LIBCLANG_PATH=C:\\Program Files\\LLVM\\bin" >> "$GITHUB_ENV"
+          else
+            choco install llvm --no-progress -y
+            echo "LIBCLANG_PATH=C:\\Program Files\\LLVM\\bin" >> "$GITHUB_ENV"
+          fi
+
       - name: Build and release (tauri-action)
         uses: tauri-apps/tauri-action@84b9d35b5fc46c1e45415bdb6144030364f7ebc5 # v0
         env:

templates/clips/desktop/README.md

@@ -6,6 +6,37 @@ A small Tauri 2.x menu-bar app that lives in the macOS menu bar / Windows system
 - **Recent** — your three most recent recordings
 - Quick links to **Open library** and **Settings**
 
+## Meeting transcription (macOS + Windows)
+
+When the calendar watcher detects a meeting, the popover surfaces a "Start
+notes" notification. Accepting it captures **both** the microphone and the
+system (speaker) audio, transcribes them locally with Whisper (whisper.cpp, no
+cloud round-trip), and streams a live mic/system-labeled transcript.
+
+Audio capture is platform-dispatched in `src-tauri/src/capture/`:
+
+- **macOS** — AVAudioEngine with VoiceProcessingIO acoustic echo cancellation
+  (mic) and ScreenCaptureKit (system audio).
+- **Windows** — [`cpal`](https://crates.io/crates/cpal): the default WASAPI
+  input device (mic) and WASAPI **loopback** of the default output device
+  (system audio). Loopback needs no OS permission prompt.
+
+The Whisper engine (`src-tauri/src/whisper_speech.rs`) and the meeting
+detection, notification, and transcript-rendering flows are otherwise identical
+across platforms.
+
+### Known Windows limitations (v1)
+
+- **No mic echo cancellation.** macOS applies hardware AEC so the mic stream
+  doesn't echo the system audio. `cpal` delivers the raw mic, so expect some
+  speaker bleed into the mic transcript when not on headphones. Mic and system
+  are transcribed and labeled separately, so this is cosmetic.
+- **No sleep / call-ended auto-stop.** The macOS sleep and call-ended watchers
+  are no-ops on Windows; the silence-based auto-stop still works. Stop notes
+  manually or via the silence timeout.
+- macOS-only features stay macOS-only: screen/window video recording, EventKit
+  local calendar, and Accessibility-based personal-vocabulary auto-learn.
+
 ## Develop
 
 First install the desktop workspace's own deps (this folder is outside the monorepo's `templates/*` glob because it ships its own Tauri/Vite toolchain):

templates/clips/desktop/src-tauri/Cargo.toml

@@ -112,13 +112,27 @@ screencapturekit = { version = "2.0.0", features = ["macos_15_0"] }
 # AudioBufferList type so we can hand SCK's CMSampleBuffer audio bytes to
 # AVAudioPCMBuffer for the speech recognizer.
 objc2-core-audio-types = "0.3"
-whisper-rs = { version = "0.16.0" }
 # NOTE: We hand-roll the few EventKit selectors we need via raw `objc2`
 # msg_send! in `eventkit.rs` rather than depending on the `objc2-event-kit`
 # crate. Its API surface drifts between minor 0.3.x releases (and not
 # every version exposes `EKEventStore::eventsMatchingPredicate`); the
 # slice we need is small enough that hand-rolling is more durable.
 
+# Windows-only: cross-platform audio capture. Covers both the microphone
+# (default WASAPI input) and system audio (WASAPI loopback — an input stream on
+# the default output device). One crate for both meeting capture streams.
+# NOTE: pin verified against crates.io before release; network was unavailable
+# in the dev environment where this was added.
+[target."cfg(target_os = \"windows\")".dependencies]
+cpal = "0.15"
+
+# Local meeting transcription via whisper.cpp. Cross-platform (bundles its own
+# native C/C++ build via CMake), needed only on the two platforms that run the
+# Whisper meeting engine. Scoped here rather than the macOS-only block (it is
+# not macOS-specific) but kept off Linux, which has no meeting capture.
+[target."cfg(any(target_os = \"macos\", target_os = \"windows\"))".dependencies]
+whisper-rs = { version = "0.16.0" }
+
 [profile.release]
 panic = "abort"
 codegen-units = 1

templates/clips/desktop/src-tauri/src/capture/macos.rs

@@ -0,0 +1,10 @@
+//! macOS capture backend — thin re-exports of the existing native impls.
+//!
+//! The high-quality macOS path is left entirely untouched: the microphone runs
+//! through `native_speech` (AVAudioEngine + VoiceProcessingIO AEC) and the
+//! system audio through `system_audio` (ScreenCaptureKit). This file only
+//! surfaces them under the platform-agnostic names the `capture` contract and
+//! `whisper_speech.rs` expect.
+
+pub(crate) use crate::native_speech::macos::{start_raw_mic_capture, RawMicCapture};
+pub(crate) use crate::system_audio::macos::{start_raw_system_capture, RawSystemCapture};

templates/clips/desktop/src-tauri/src/capture/mod.rs

@@ -0,0 +1,39 @@
+//! Platform-dispatched audio capture for meeting transcription.
+//!
+//! The local Whisper engine (`whisper_speech.rs`) needs exactly two capture
+//! primitives — a microphone stream and a system-audio (loopback) stream —
+//! each forwarding mono `f32` samples to a callback and exposing the hardware
+//! sample rate plus a `stop()`. Everything else in the meeting pipeline
+//! (detection, notifications, transcript rendering) is already cross-platform.
+//!
+//! This module owns that public contract and dispatches to the right backend
+//! at compile time:
+//!
+//!   - macOS  → thin re-exports of the proven `native_speech` (AVAudioEngine +
+//!              VPIO AEC) mic path and `system_audio` (ScreenCaptureKit) loopback.
+//!   - Windows → `cpal` mic + WASAPI loopback (see `windows.rs`).
+//!
+//! Both backends expose the same names so `whisper_speech.rs` stays
+//! platform-agnostic:
+//!
+//! ```ignore
+//! start_raw_mic_capture(app, mic_device_id, mic_device_label, on_samples) -> RawMicCapture
+//! start_raw_system_capture(app, on_samples) -> RawSystemCapture
+//! ```
+//!
+//! Each handle type exposes `sample_rate() -> f64` and `stop()` so the session
+//! teardown in `whisper_speech.rs` works unchanged across platforms.
+
+#[cfg(target_os = "macos")]
+mod macos;
+#[cfg(target_os = "macos")]
+pub(crate) use macos::{
+    start_raw_mic_capture, start_raw_system_capture, RawMicCapture, RawSystemCapture,
+};
+
+#[cfg(target_os = "windows")]
+mod windows;
+#[cfg(target_os = "windows")]
+pub(crate) use windows::{
+    start_raw_mic_capture, start_raw_system_capture, RawMicCapture, RawSystemCapture,
+};