Moonshine speech engine + Wayland overlay fixes (v0.3.0)

.github/workflows/release.yml

@@ -253,7 +253,9 @@ jobs:
         env:
           APPIMAGE_EXTRACT_AND_RUN: 1
           QT_QPA_PLATFORM: offscreen
-        run: npm run tauri build
+        # `moonshine` ships the Moonshine ONNX engine alongside whisper-cpp in
+        # every release build (links ONNX Runtime, fetched at build time).
+        run: npm run tauri build -- --features moonshine
 
       # Portable GPU AppImage. whisper.cpp's Vulkan backend links only
       # libvulkan (lightweight, unlike CUDA), so linuxdeploy bundles it fine.
@@ -262,7 +264,7 @@ jobs:
         env:
           APPIMAGE_EXTRACT_AND_RUN: 1
           QT_QPA_PLATFORM: offscreen
-        run: npm run tauri build -- --features vulkan
+        run: npm run tauri build -- --features vulkan,moonshine
 
       # linuxdeploy bundles whatever graphics/Wayland/input libraries exist on
       # the build host. On ubuntu-22.04 those are older than modern desktops, and
@@ -312,7 +314,7 @@ jobs:
           QT_QPA_PLATFORM: offscreen
         # Call tauri directly so --bundles lands before the cargo -- separator;
         # `npm run tauri build -- --bundles deb` wrongly sends --bundles to cargo.
-        run: node_modules/.bin/tauri build --bundles deb -- --features cuda
+        run: node_modules/.bin/tauri build --bundles deb -- --features cuda,moonshine
 
       # ── Tauri build (Windows) ───────────────────────────────────────────────
       # NSIS (setup.exe) only — the advertised Windows installer. The WiX/MSI
@@ -321,7 +323,7 @@ jobs:
       # `--bundles nsis` reaches tauri instead of cargo.
       - name: Build installer (Windows, CPU)
         if: runner.os == 'Windows' && !matrix.cuda
-        run: npx tauri build --bundles nsis
+        run: npx tauri build --bundles nsis -- --features moonshine
 
       - name: Copy real CUDA DLLs (Windows, CUDA)
         if: runner.os == 'Windows' && matrix.cuda
@@ -333,7 +335,7 @@ jobs:
 
       - name: Build installer (Windows, CUDA)
         if: runner.os == 'Windows' && matrix.cuda
-        run: npx tauri build --bundles nsis -- --features cuda
+        run: npx tauri build --bundles nsis -- --features cuda,moonshine
 
       # ── Collect and rename Linux artifacts ──────────────────────────────────
       # Search the whole workspace so the path is correct regardless of whether

Cargo.toml

@@ -15,7 +15,7 @@ members = [
 resolver = "2"
 
 [workspace.package]
-version = "0.2.8"
+version = "0.3.0"
 edition = "2021"
 authors = ["VoxCtrl Contributors"]
 license = "MIT"

build_appimage.sh

@@ -207,12 +207,15 @@ if [ "$CUDA_FOUND" = false ] && [ "$HAS_NVIDIA_GPU" = true ]; then
 fi
 
 info "Running Tauri release compiler with headless PATH and CUDA injection..."
+# The Moonshine ONNX backend is always compiled in so both speech engines
+# (whisper-cpp and Moonshine) are selectable in every AppImage. It links ONNX
+# Runtime, fetched at build time, so this step needs network access.
 if [ "$CUDA_FOUND" = true ]; then
-    info "CUDA detected. Compiling with GPU support..."
-    npx tauri build -- --features cuda
+    info "CUDA detected. Compiling with GPU support (whisper-cpp + Moonshine)..."
+    npx tauri build -- --features cuda,moonshine
 else
-    info "CUDA not detected. Compiling for CPU only..."
-    npx tauri build
+    info "CUDA not detected. Compiling for CPU only (whisper-cpp + Moonshine)..."
+    npx tauri build -- --features moonshine
 fi
 
 ok "Compilation finished successfully."

crates/voxctrl-inference/Cargo.toml

@@ -30,15 +30,28 @@ cuda = ["whisper-rs/cuda"]
 # CUDA's heavy runtime. Falls back to CPU when no Vulkan device is present.
 vulkan = ["whisper-rs/vulkan"]
 
-# Moonshine ONNX backend via onnxruntime
-moonshine = ["dep:ort", "dep:ndarray"]
+# Moonshine ONNX backend via onnxruntime. `ort` keeps its default features,
+# which include `download-binaries`: ONNX Runtime is fetched at build time and
+# linked into the binary, so the shipped app is self-contained and needs no
+# system `libonnxruntime` at runtime (nothing else in the workspace provides
+# one). Decoding uses the `tokenizers` crate with a tokenizer embedded in the
+# binary (see `assets/moonshine_tokenizer.json`).
+moonshine = ["dep:ort", "dep:tokenizers"]
 
 [dependencies.ort]
 version = "2.0.0-rc.12"
 optional = true
-features = ["download-binaries"]
+# Default features include `download-binaries`, which fetches a prebuilt ONNX
+# Runtime at build time and links it statically, so the resulting binary is
+# self-contained (no system libonnxruntime needed at runtime). Building the
+# `moonshine` feature therefore requires network access to the ONNX Runtime
+# binary host at compile time. To build fully offline instead, switch to
+# `default-features = false, features = ["load-dynamic", "std", "api-24"]` and
+# ship a `libonnxruntime` alongside the app.
 
-[dependencies.ndarray]
-version = "0.16"
+[dependencies.tokenizers]
+version = "0.21"
 optional = true
+default-features = false
+features = ["onig"]