malithedeveloper · malithedeveloper · Aug 14, 2026 · Aug 14, 2026
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
@@ -74,15 +74,15 @@ jobs:
           cmake
           -DVOXLOCAL_PLATFORM=linux
           -DVOXLOCAL_PAYLOAD_DIR=${{ github.workspace }}/package/plugin
-          -DVOXLOCAL_OUTPUT=${{ github.workspace }}/release/VoxLocal-Setup-1.1.1-Linux-x86_64.run
+          -DVOXLOCAL_OUTPUT=${{ github.workspace }}/release/VoxLocal-Setup-1.1.2-Linux-x86_64.run
           -DVOXLOCAL_WORK_DIR=${{ github.workspace }}/build/installer-linux
           -P cmake/package-installer.cmake
       - name: Mark setup executable
-        run: chmod +x release/VoxLocal-Setup-1.1.1-Linux-x86_64.run
+        run: chmod +x release/VoxLocal-Setup-1.1.2-Linux-x86_64.run
       - uses: actions/upload-artifact@v4
         with:
-          name: VoxLocal-Setup-1.1.1-Linux-x86_64
-          path: release/VoxLocal-Setup-1.1.1-Linux-x86_64.run
+          name: VoxLocal-Setup-1.1.2-Linux-x86_64
+          path: release/VoxLocal-Setup-1.1.2-Linux-x86_64.run
           if-no-files-found: error
 
   windows-plugin:
@@ -170,13 +170,13 @@ jobs:
           cmake
           -DVOXLOCAL_PLATFORM=windows
           -DVOXLOCAL_PAYLOAD_DIR=${{ github.workspace }}/package/plugin
-          -DVOXLOCAL_OUTPUT=${{ github.workspace }}/release/VoxLocal-Setup-1.1.1-Windows-x64.exe
+          -DVOXLOCAL_OUTPUT=${{ github.workspace }}/release/VoxLocal-Setup-1.1.2-Windows-x64.exe
           -DVOXLOCAL_WORK_DIR=${{ github.workspace }}/build/installer-windows
           -P cmake/package-installer.cmake
       - uses: actions/upload-artifact@v4
         with:
-          name: VoxLocal-Setup-1.1.1-Windows-x64
-          path: release/VoxLocal-Setup-1.1.1-Windows-x64.exe
+          name: VoxLocal-Setup-1.1.2-Windows-x64
+          path: release/VoxLocal-Setup-1.1.2-Windows-x64.exe
           if-no-files-found: error
 
   macos-plugin:
@@ -281,13 +281,13 @@ jobs:
           cmake
           -DVOXLOCAL_PLATFORM=macos
           -DVOXLOCAL_PAYLOAD_DIR=${{ github.workspace }}/package/plugin
-          -DVOXLOCAL_OUTPUT=${{ github.workspace }}/release/VoxLocal-Setup-1.1.1-macOS-universal.dmg
+          -DVOXLOCAL_OUTPUT=${{ github.workspace }}/release/VoxLocal-Setup-1.1.2-macOS-universal.dmg
           -DVOXLOCAL_WORK_DIR=${{ github.workspace }}/build/installer-macos
           -P cmake/package-installer.cmake
       - uses: actions/upload-artifact@v4
         with:
-          name: VoxLocal-Setup-1.1.1-macOS-universal
-          path: release/VoxLocal-Setup-1.1.1-macOS-universal.dmg
+          name: VoxLocal-Setup-1.1.2-macOS-universal
+          path: release/VoxLocal-Setup-1.1.2-macOS-universal.dmg
           if-no-files-found: error
 
   release:
@@ -301,11 +301,11 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/download-artifact@v4
         with:
-          pattern: VoxLocal-Setup-1.1.1-*
+          pattern: VoxLocal-Setup-1.1.2-*
           path: dist
           merge-multiple: true
       - uses: softprops/action-gh-release@v2
         with:
-          body_path: docs/RELEASE_NOTES_1.1.1.md
+          body_path: docs/RELEASE_NOTES_1.1.2.md
           files: dist/*
           fail_on_unmatched_files: true
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,17 @@
 # Changelog
 
+## 1.1.2 — 2026-08-14
+
+- Fixed Windows model startup by applying DirectML's required sequential/no-memory-pattern session settings.
+- Updated Windows to Microsoft ONNX Runtime DirectML 1.24.4 and added automatic strongest-GPU selection with full CPU fallback.
+- Moved model initialization to a background worker after OBS finishes loading, preventing the OBS interface from freezing during model startup.
+- Added `Ask at startup`, `Load automatically`, and `Do not load` model-memory policies to both Welcome and the VoxLocal dock; the default is `Ask at startup`.
+- Kept the VoxLocal dock hidden until Welcome finishes and made Welcome non-dismissible, while adding a `Skip` button to every page so optional setup sections never block progress.
+- Added a manual `Load model now` action and a visible loaded backend indicator to the dock.
+- Logged model initialization and inference failures to the OBS log instead of showing them only in the dock.
+- Prevented harmless reads from an already-closed TLS reply after a completed model download.
+- Made the Windows setup remove VoxLocal's obsolete `%APPDATA%` plugin copy when installing to `%ProgramData%`, avoiding duplicate-version conflicts.
+
 ## 1.1.1 — 2026-08-13
 
 - Updated the Windows and macOS SDK builds to OBS Studio 32.2.1 and its matching 2026-07-15 dependencies.

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.28)
 
-project(VoxLocal VERSION 1.1.1 LANGUAGES CXX)
+project(VoxLocal VERSION 1.1.2 LANGUAGES CXX)
 
 set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
@@ -87,6 +87,7 @@ if(VOXLOCAL_ENABLE_ONNX)
     target_compile_definitions(voxlocal-core PRIVATE VOXLOCAL_HAVE_ONNXRUNTIME=1)
     if(VOXLOCAL_ORT_DIRECTML)
       target_compile_definitions(voxlocal-core PRIVATE VOXLOCAL_ORT_DIRECTML=1)
+      target_link_libraries(voxlocal-core PRIVATE dxgi)
     endif()
   else()
     message(STATUS "ONNX Runtime not found: VoxLocal will build with model-management support but inference disabled")

diff --git a/README.md b/README.md
@@ -11,6 +11,8 @@ VoxLocal is a native OBS Studio plugin that turns live Kick chat commands into l
 ## Features
 
 - Local Chatterbox Multilingual ONNX inference with Turkish and 22 other languages, including safe handling for tokenizer characters that are absent from the exported embedding table.
+- Non-blocking model initialization after OBS is ready, with `Ask at startup`, `Load automatically`, and `Do not load` memory policies.
+- Automatic hardware backend selection with DirectML GPU/CPU fallback on Windows, CoreML detection on macOS, and CUDA detection when a CUDA-enabled Linux runtime is supplied.
 - Checksum-verified model download with pause/resume support and detailed byte progress.
 - Zero-shot voice cloning from audio or video. FFmpeg extracts the first audio track and stores a validated 24 kHz mono WAV locally.
 - Multiple personas, each with its own chat command, language mode, voice sample, and viewer role policy.
@@ -45,15 +47,15 @@ VoxLocal attempts DirectML on Windows, CoreML on macOS, and CUDA on Linux when t
 
 Close OBS, download the one setup matching your platform from [Releases](../../releases), and run it:
 
-- `VoxLocal-Setup-1.1.1-Windows-x64.exe`
-- `VoxLocal-Setup-1.1.1-macOS-universal.dmg` for Apple Silicon and Intel Macs
-- `VoxLocal-Setup-1.1.1-Linux-x86_64.run`
+- `VoxLocal-Setup-1.1.2-Windows-x64.exe`
+- `VoxLocal-Setup-1.1.2-macOS-universal.dmg` for Apple Silicon and Intel Macs
+- `VoxLocal-Setup-1.1.2-Linux-x86_64.run`
 
 On Linux, make the downloaded setup executable first if your browser removed its executable bit:
 
 ```bash
-chmod +x VoxLocal-Setup-1.1.1-Linux-x86_64.run
-./VoxLocal-Setup-1.1.1-Linux-x86_64.run
+chmod +x VoxLocal-Setup-1.1.2-Linux-x86_64.run
+./VoxLocal-Setup-1.1.2-Linux-x86_64.run
 ```
 
 The setup detects and selects the plugin folder scanned by OBS Studio 32.2.1 automatically:
@@ -69,10 +71,13 @@ Windows and macOS setups are currently unsigned. Windows SmartScreen or macOS Ga
 ## First run
 
 1. Select English or Turkish.
-2. Download and verify the local model. Setup cannot continue before it finishes.
-3. Enter the public Kick channel slug.
-4. Create a persona with a name, command, language, access policy, and audio or video sample.
-5. VoxLocal adds **VoxLocal Overlay** to the current scene.
+2. Download and verify the local model, or use **Skip** to do it later.
+3. Choose whether the model should ask, load automatically, or remain unloaded when OBS starts.
+4. Enter the public Kick channel slug or skip it.
+5. Create a persona with a name, command, language, access policy, and audio or video sample, or skip it.
+6. VoxLocal adds **VoxLocal Overlay** to the current scene.
+
+The VoxLocal dock remains hidden until Welcome finishes. Welcome cannot be accidentally dismissed, but each page has a **Skip** button and all sections can be completed later from the dock. Model weights are loaded into memory on a background worker only after OBS has finished opening. The dock also provides **Load model now** and shows the selected CPU/GPU backend.
 
 The settings dock opens automatically. Reopen it from **Tools → VoxLocal Settings** or press `Ctrl+Shift+V`.
 

diff --git a/buildspec.json b/buildspec.json
@@ -38,7 +38,7 @@
   },
   "name": "voxlocal",
   "displayName": "VoxLocal",
-  "version": "1.1.1",
+  "version": "1.1.2",
   "author": "malidev",
   "website": "https://github.com/malithedeveloper/voxlocal",
   "email": ""

diff --git a/cmake/onnxruntime.cmake b/cmake/onnxruntime.cmake
@@ -2,8 +2,8 @@ include(FetchContent)
 
 function(voxlocal_fetch_onnxruntime)
   if(WIN32)
-    set(_ort_url "https://github.com/microsoft/onnxruntime/releases/download/v1.22.1/Microsoft.ML.OnnxRuntime.DirectML.1.22.1.nupkg")
-    set(_ort_hash "9fe1d91a697fb8a68abd29fdc2a88dea8f864ee74e758a055ce63c9e023c6908")
+    set(_ort_url "https://api.nuget.org/v3-flatcontainer/microsoft.ml.onnxruntime.directml/1.24.4/microsoft.ml.onnxruntime.directml.1.24.4.nupkg")
+    set(_ort_hash "57e9f11b73437bef7a309496135d4c1f96b1a8e9ddba60013fa27bfc1d788681")
     set(_ort_include_suffix "build/native/include")
     set(_ort_library_suffix "runtimes/win-x64/native")
     set(_ort_directml ON)

diff --git a/cmake/package-installer.cmake b/cmake/package-installer.cmake
@@ -43,7 +43,7 @@ endif()
 get_filename_component(VOXLOCAL_SOURCE_DIR "${CMAKE_CURRENT_LIST_DIR}/.." ABSOLUTE)
 file(READ "${VOXLOCAL_SOURCE_DIR}/buildspec.json" buildspec)
 string(JSON VOXLOCAL_VERSION GET "${buildspec}" version)
-set(VOXLOCAL_RELEASE_DATE "2026-08-13")
+set(VOXLOCAL_RELEASE_DATE "2026-08-14")
 
 if(VOXLOCAL_PLATFORM STREQUAL "windows")
   set(VOXLOCAL_DEFAULT_TARGET "@RootDir@/ProgramData/obs-studio/plugins")
@@ -76,6 +76,10 @@ configure_file(
   "${VOXLOCAL_SOURCE_DIR}/installer/packages/dev.malithedeveloper.voxlocal/meta/package.xml.in"
   "${meta_dir}/package.xml"
   @ONLY)
+configure_file(
+  "${VOXLOCAL_SOURCE_DIR}/installer/packages/dev.malithedeveloper.voxlocal/meta/installscript.qs"
+  "${meta_dir}/installscript.qs"
+  COPYONLY)
 configure_file(
   "${VOXLOCAL_SOURCE_DIR}/installer/config/controller.qs"
   "${config_dir}/controller.qs"

diff --git a/docs/RELEASE_NOTES_1.1.2.md b/docs/RELEASE_NOTES_1.1.2.md
@@ -0,0 +1,34 @@
+# VoxLocal 1.1.2
+
+VoxLocal 1.1.2 fixes Windows TTS model startup, makes model loading non-blocking, and gives the streamer explicit control over whether the model enters memory when OBS starts.
+
+## Download
+
+Choose one setup for your operating system:
+
+- `VoxLocal-Setup-1.1.2-Windows-x64.exe`
+- `VoxLocal-Setup-1.1.2-macOS-universal.dmg` for Apple Silicon and Intel Macs
+- `VoxLocal-Setup-1.1.2-Linux-x86_64.run`
+
+Close OBS before running the setup. The Windows setup uses `%ProgramData%\obs-studio\plugins` and removes VoxLocal's obsolete `%APPDATA%` plugin copy. Existing settings, voices, and downloaded models remain untouched.
+
+## Changes
+
+- Applied the session options required by ONNX Runtime DirectML, fixing model initialization failures on Windows.
+- Updated the Windows inference runtime from DirectML 1.22.1 to 1.24.4.
+- Automatically chooses the hardware GPU with the most dedicated video memory on Windows. If the GPU or any model session cannot initialize, VoxLocal recreates every session on CPU automatically.
+- Detects the available ONNX execution provider on each platform and displays the backend selected for the current session.
+- Opens the OBS interface first and initializes the model on a background worker afterward, so loading model weights does not block OBS startup.
+- Added three model startup behaviors to Welcome and the dock: **Ask at startup** (default), **Load automatically**, and **Do not load**.
+- Added **Load model now** to the dock for manual per-session loading.
+- Keeps the VoxLocal dock closed while Welcome is active. Welcome cannot be dismissed with Cancel, Escape, or the window close button, but every page has **Skip** and all setup sections are optional.
+- Writes model load and inference failures into the OBS log for useful diagnostics.
+- Avoids reading a closed TLS socket after a completed model download.
+- Retains the 1.1.1 Turkish tokenizer bounds fix and larger Minimal overlay text.
+
+## Notes
+
+- The startup behavior controls loading the already-downloaded model into memory; it does not delete or redownload model files.
+- Windows and macOS setups are unsigned and may require SmartScreen or Gatekeeper confirmation.
+- The Linux setup targets native x86-64 OBS installations. Flatpak OBS uses Flatpak's own plugin mechanism.
+- FFmpeg must be available locally to import voice samples.
diff --git a/include/voxlocal/runtime.hpp b/include/voxlocal/runtime.hpp
@@ -29,19 +29,23 @@ class VoxLocalRuntime final : public QObject
   [[nodiscard]] KickConnector *kickConnector() { return &kick_; }
   [[nodiscard]] QString status() const { return status_; }
   [[nodiscard]] bool engineReady() const { return engine_->isReady(); }
+  [[nodiscard]] bool engineLoading() const { return engineLoading_; }
+  [[nodiscard]] QString engineBackend() const { return engine_->backendName(); }
   [[nodiscard]] bool isVoiceImporting(const QString &personaId = {}) const;
 
   bool applySettings(Settings settings, QString *error = nullptr);
   QString importVoice(const QString &sourcePath, const QString &personaId, QString *error = nullptr) const;
   void importVoiceAsync(const QString &sourcePath, const QString &personaId);
   void start();
   void stop();
+  void loadModel();
   bool preview(const QString &personaId, const QString &text, QString *error = nullptr);
 
 signals:
   void settingsChanged();
   void statusChanged(const QString &status);
   void modelProgress(qint64 received, qint64 total, const QString &fileName);
+  void engineLoadingChanged(bool loading);
   void voiceImportProgress(const QString &personaId, int percent, const QString &operation);
   void voiceImportFinished(const QString &personaId, const QString &voicePath);
   void voiceImportFailed(const QString &personaId, const QString &error);
@@ -68,10 +72,12 @@ class VoxLocalRuntime final : public QObject
   std::shared_ptr<ChatterboxEngine> engine_;
   SpeechQueue queue_;
   QThreadPool voiceImportPool_;
+  QThreadPool modelLoadPool_;
   QSet<QString> voiceImports_;
   QString activeChatRequestId_;
   QDateTime globalCooldownUntil_;
   bool chatGenerationActive_ = false;
+  bool engineLoading_ = false;
   QString status_ = QStringLiteral("idle");
 };
 

diff --git a/include/voxlocal/speech-queue.hpp b/include/voxlocal/speech-queue.hpp
@@ -20,6 +20,7 @@ class SpeechQueue final : public QObject
   ~SpeechQueue() override;
 
   void setCapacity(int capacity);
+  bool setEngine(std::shared_ptr<ITtsEngine> engine);
   [[nodiscard]] int size() const;
   bool enqueue(const QueueItem &item);
 

diff --git a/include/voxlocal/types.hpp b/include/voxlocal/types.hpp
@@ -15,13 +15,14 @@
 
 namespace voxlocal {
 
-inline constexpr int kConfigSchemaVersion = 3;
+inline constexpr int kConfigSchemaVersion = 4;
 inline constexpr int kDefaultMaxTextLength = 250;
 inline constexpr int kDefaultGlobalCooldownSeconds = 10;
 inline constexpr int kDefaultQueueCapacity = 10;
 
 enum class InterfaceLanguage { English, Turkish };
 enum class LanguageMode { Fixed, Automatic };
+enum class ModelStartupBehavior { Ask, AlwaysLoad, NeverLoad };
 enum class UserRole : std::uint32_t {
   None = 0,
   Everyone = 1u << 0,
@@ -112,6 +113,7 @@ struct Settings {
   InterfaceLanguage interfaceLanguage = InterfaceLanguage::English;
   QString defaultSpeechLanguage = QStringLiteral("en");
   bool welcomeCompleted = false;
+  ModelStartupBehavior modelStartupBehavior = ModelStartupBehavior::Ask;
   bool readUrls = false;
   bool ttsEnabled = true;
   int globalCooldownSeconds = kDefaultGlobalCooldownSeconds;

diff --git a/include/voxlocal/voxlocal-dock.hpp b/include/voxlocal/voxlocal-dock.hpp
@@ -42,6 +42,7 @@ class VoxLocalDock final : public QWidget
   void chooseColor(QPushButton *button, QString *target);
   void updateModelProgress(qint64 received, qint64 total, const QString &fileName);
   void updateModelDownloadState(bool installing);
+  void updateModelRuntimeState();
   void updateVoiceImportProgress(const QString &personaId, int percent, const QString &operation);
   void finishVoiceImport(const QString &personaId, const QString &voicePath);
   void failVoiceImport(const QString &personaId, const QString &error);
@@ -59,6 +60,9 @@ class VoxLocalDock final : public QWidget
   QLabel *model_ = nullptr;
   QProgressBar *modelProgress_ = nullptr;
   QPushButton *modelInstall_ = nullptr;
+  QPushButton *modelLoad_ = nullptr;
+  QLabel *modelRuntime_ = nullptr;
+  QComboBox *modelStartup_ = nullptr;
   QComboBox *interfaceLanguage_ = nullptr;
   QLineEdit *channel_ = nullptr;
   QListWidget *personas_ = nullptr;

diff --git a/include/voxlocal/welcome-wizard.hpp b/include/voxlocal/welcome-wizard.hpp
@@ -6,6 +6,7 @@
 #include <QWizard>
 
 class QComboBox;
+class QCloseEvent;
 class QLabel;
 class QLineEdit;
 class QListWidget;
@@ -20,9 +21,12 @@ class WelcomeWizard final : public QWizard
 
 public:
   explicit WelcomeWizard(VoxLocalRuntime *runtime, QWidget *parent = nullptr);
+  void closeForShutdown();
 
 protected:
   void accept() override;
+  void reject() override;
+  void closeEvent(QCloseEvent *event) override;
 
 private:
   struct LocalizedWidget {
@@ -42,6 +46,7 @@ class WelcomeWizard final : public QWizard
   VoxLocalRuntime *runtime_;
   QVector<LocalizedWidget> localizedWidgets_;
   QComboBox *interfaceLanguage_ = nullptr;
+  QComboBox *modelStartup_ = nullptr;
   QLabel *modelInfo_ = nullptr;
   QLabel *modelStatus_ = nullptr;
   QProgressBar *modelProgress_ = nullptr;
@@ -59,6 +64,7 @@ class WelcomeWizard final : public QWizard
   QLineEdit *voice_ = nullptr;
   QComboBox *language_ = nullptr;
   QComboBox *access_ = nullptr;
+  bool allowClose_ = false;
 };
 
 } // namespace voxlocal
diff --git a/installer/packages/dev.malithedeveloper.voxlocal/meta/installscript.qs b/installer/packages/dev.malithedeveloper.voxlocal/meta/installscript.qs
@@ -0,0 +1,26 @@
+function Component()
+{
+}
+
+Component.prototype.createOperations = function()
+{
+    component.createOperations();
+
+    if (systemInfo.productType !== "windows")
+        return;
+
+    var appData = installer.environmentVariable("APPDATA");
+    if (appData === "")
+        return;
+
+    var stalePlugin = appData + "/obs-studio/plugins/voxlocal";
+    var selectedTarget = installer.value("TargetDir").replace(/\\/g, "/").toLowerCase();
+    var staleParent = (appData + "/obs-studio/plugins").replace(/\\/g, "/").toLowerCase();
+    if (selectedTarget === staleParent)
+        return;
+
+    var systemRoot = installer.environmentVariable("SystemRoot");
+    var command = (systemRoot !== "" ? systemRoot : "C:/Windows") + "/System32/cmd.exe";
+    component.addOperation("Execute", command, "/D", "/S", "/C",
+                           "if exist \"" + stalePlugin + "\" rmdir /S /Q \"" + stalePlugin + "\"");
+}
diff --git a/installer/packages/dev.malithedeveloper.voxlocal/meta/package.xml.in b/installer/packages/dev.malithedeveloper.voxlocal/meta/package.xml.in
@@ -6,6 +6,7 @@
     <ReleaseDate>@VOXLOCAL_RELEASE_DATE@</ReleaseDate>
     <Name>dev.malithedeveloper.voxlocal</Name>
     <Default>true</Default>
+    <Script>installscript.qs</Script>
 @VOXLOCAL_ADMIN_REQUIREMENT@
     <Licenses>
         <License name="GNU GPL 2.0 or later" file="LICENSE" />