feat: add lightweight real-time native voice preview

.github/workflows/native-preview.yml

@@ -0,0 +1,46 @@
+name: Native Preview
+
+on:
+  workflow_dispatch:
+  pull_request:
+    paths:
+      - 'native-app/**'
+      - 'catalog/models.json'
+      - '.github/workflows/native-preview.yml'
+  push:
+    branches: [main]
+    paths:
+      - 'native-app/**'
+      - 'catalog/models.json'
+      - '.github/workflows/native-preview.yml'
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: windows-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v7
+
+      - name: Setup .NET 10
+        uses: actions/setup-dotnet@v6
+        with:
+          dotnet-version: '10.0.x'
+
+      - name: Build
+        shell: pwsh
+        run: dotnet build ./native-app/VoxType.Native.slnx -c Release
+
+      - name: Package framework-dependent preview
+        shell: pwsh
+        run: pwsh -NoProfile -File ./native-app/build.ps1 -SkipBuild
+
+      - name: Upload preview zip
+        uses: actions/upload-artifact@v7
+        with:
+          name: VoxType-NativePreview-win-x64
+          path: native-app/publish/VoxType.NativePreview-win-x64.zip
+          if-no-files-found: error
+          retention-days: 14

.gitignore

@@ -20,6 +20,11 @@ app/src-tauri/bundle-resources/runtime/voxtype-runtime/*
 !app/src-tauri/bundle-resources/runtime/voxtype-runtime/.gitkeep
 app/src-tauri/bundle-resources/catalog/models.json
 
+# Native .NET preview
+native-app/**/bin/
+native-app/**/obj/
+native-app/publish/
+
 # Plugin
 plugin/bin/
 plugin/obj/

AGENTS.md

@@ -13,6 +13,7 @@ VoxType 是 Windows 本地语音输入应用，也是 Quicker 语音输入扩展
 | 目录 | 职责 |
 |------|------|
 | `app/` | Tauri 桌面宿主：热键、录音、浮层、设置、文本提交 |
+| `native-app/` | .NET 10/WPF 原生体验版与可复用识别 core；不依赖 Python runtime |
 | `runtime/` | 独立 ASR sidecar：模型加载、识别、WebSocket 会话 |
 | `plugin/` | Quicker 兼容适配器；不承载识别引擎 |
 | `catalog/` | 随安装包发布的模型元数据；模型权重不入库 |
@@ -44,6 +45,11 @@ pnpm install
 pnpm build
 cargo test --manifest-path src-tauri/Cargo.toml
 
+# Native .NET preview
+cd ../../native-app
+dotnet build VoxType.Native.slnx -c Release
+pwsh ./build.ps1 -SkipBuild
+
 # 模型目录
 cd ..
 pwsh ./scripts/validate-models-catalog.ps1

README.md

@@ -12,6 +12,7 @@ ASR runtime uses [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) (SenseVoic
 | `app/` | Tauri client — settings, overlay, hotkeys, direct typing, Quicker HTTP API |
 | `plugin/` | Quicker plugin — button triggers dictation |
 | `catalog/models.json` | Public model catalog for settings UI |
+| `native-app/` | .NET 10/WPF native preview — no Python sidecar |
 
 ## Quick start (dev)
 
@@ -35,6 +36,8 @@ pnpm tauri dev
 
 Data directory: `%LOCALAPPDATA%\VoxType\` (models, settings).
 
+For the smaller native preview intended to validate future `QuickerPc/Quicker` integration, see [`native-app/`](native-app/README.md).
+
 ## Model download
 
 Model **weights** are not bundled (too large). Model **download URLs** are written in `catalog/models.json` and ship inside the installer at `catalog/models.json` — the app reads this file locally (no remote catalog fetch). Users download weights from the settings UI on first use.

docs/quicker-integration-research.md

@@ -17,6 +17,7 @@ Windows Text Services Framework（TSF）适合长期追求输入法级体验，
 - sherpa-onnx + SenseVoice 已能离线工作，runtime 目录也预留了其它识别后端。
 - 当前桌面端直接模拟 Unicode 键入；遇到管理员窗口、特殊编辑器、终端或复杂输入法状态时兼容性有限。
 - 当前 Quicker 插件经 `127.0.0.1:6020` 调用 VoxType，只是外部控制桥，不是 Quicker 原生 provider。
+- `native-app/` 已提供 `.NET 10 + WPF + sherpa-onnx C#` 的无 Python 体验版，并用约 25 MB 的中文在线 Zipformer 验证 40 ms 音频分块、partial/final 与端点重置；它不改变“Quicker 主进程不直接承载模型”的隔离边界。
 
 ### QuickerPc/Quicker
 
@@ -125,6 +126,7 @@ v2 在会话前增加能力握手，至少返回：
 - 抽出 JSON schema、示例音频和协议一致性测试。
 - runtime 增加 capabilities/版本信息和控制 API token。
 - 为 .NET 提供薄客户端；不把模型推理依赖带进客户端包。
+- 以 `native-app/VoxType.NativeCore` 为基础增加无 UI worker，完成稳定性与内存驻留测试后，再决定是否替换 Python runtime。
 
 ### P2：Quicker 原生 provider
 

native-app/BENCHMARKS.md

@@ -0,0 +1,46 @@
+# Native Preview 基线
+
+> 日期：2026-08-07。数据用于比较架构，不代表所有设备上的保证值。
+
+## 环境
+
+- CPU：Intel Core i9-14900HX（24 核 / 32 线程）
+- 系统：Windows x64
+- SDK / Runtime：.NET 10.0.301 / .NET Desktop Runtime 10
+- 推理：sherpa-onnx 1.13.4，在线 CTC，CPU provider，最多 4 线程
+- 模型：Zipformer 中文实时 Small INT8，26,342,340-byte ONNX
+- 音频：16 kHz mono，5.00 秒中文 WAV
+
+## 结果
+
+| 指标 | 结果 |
+|------|------|
+| 模型加载 | 1.19–2.55 s（冷热启动） |
+| 加速分块识别 | 0.14–0.20 s，RTF 0.027–0.040 |
+| 实时分块播放 | 9 次文字更新，首个 partial 约 0.57 s |
+| 发布目录 | 21.4 MB |
+| zip | 8.1 MB |
+| 模型加载后的进程 Working Set | 约 281 MB |
+| 模型加载后的 Private Memory | 约 232 MB |
+
+`v0.1.12` 的 Python/PyInstaller + Tauri NSIS 安装包为 25,001,555 bytes（约 23.8 MiB）。Native Preview 的 zip 为约 8.1 MiB，压缩体积降低约 66%；两者都不包含模型权重。流式模型文件合计约 25.1 MiB，相比原 SenseVoice 文件的 228.45 MiB 降低约 89%。
+
+## 解释
+
+- `--realtime` 按 40 ms 节奏送入 chunk；首个 partial 是用户感知延迟，比停止录音后的总 decode 时间更有代表性。
+- 初版没有为流式模型提供左侧上下文，首词在同一音频上误识别为“算好”；补充官方文件解码示例采用的 300 ms 静音后变为“上好”，快速与真实节奏的最终文本一致。
+- WPF UI、录音回调和在线 decode 分属不同线程；无界音频队列不主动丢 chunk，结束时先停止 NAudio 再刷新最终上下文。
+- 小模型明显降低磁盘和常驻内存，但本轮样例首词识别为“算好”而非“大家好”，也不自带标点。上线前必须用实际口音、噪声和 Quicker 术语集比较准确率。
+- 发布目录仍主要由 `onnxruntime.dll` 与 `sherpa-onnx-c-api.dll` 构成。后续 Quicker 接入仍建议使用独立原生 worker，隔离 native DLL 生命周期。
+
+## 复现
+
+```powershell
+pwsh ./build.ps1
+
+dotnet run --project ./src/VoxType.NativeSmoke -c Release -- `
+  <audio.wav> `
+  "$env:LOCALAPPDATA\VoxType\models\zipformer-zh-streaming-small" `
+  1 `
+  --realtime
+```

native-app/Directory.Build.props

@@ -0,0 +1,10 @@
+<Project>
+  <!-- Keep this standalone project isolated from Directory.Build.props files in parent checkouts. -->
+  <PropertyGroup>
+    <LangVersion>latest</LangVersion>
+    <TreatWarningsAsErrors>true</TreatWarningsAsErrors>
+    <DefaultItemExcludes>$(DefaultItemExcludes);publish/**;**/publish/**</DefaultItemExcludes>
+    <QkDisableQkref>true</QkDisableQkref>
+    <QkDisablePolySharp>true</QkDisablePolySharp>
+  </PropertyGroup>
+</Project>

native-app/Directory.Build.targets

@@ -0,0 +1,3 @@
+<Project>
+  <!-- Intentionally empty: stop parent repositories from injecting build targets. -->
+</Project>

native-app/Directory.Packages.props

@@ -0,0 +1,10 @@
+<Project>
+  <PropertyGroup>
+    <ManagePackageVersionsCentrally>true</ManagePackageVersionsCentrally>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <PackageVersion Include="NAudio.WinMM" Version="2.3.0" />
+    <PackageVersion Include="org.k2fsa.sherpa.onnx" Version="1.13.4" />
+  </ItemGroup>
+</Project>

native-app/README.md

@@ -0,0 +1,51 @@
+# VoxType Native Preview
+
+面向 `QuickerPc/Quicker` 的最小原生实时语音识别体验版。它使用 `.NET 10 + WPF + NAudio + sherpa-onnx C#`，直接加载中文流式 Zipformer CTC，不启动 Python、aiohttp 或 PyInstaller runtime。
+
+## 体验
+
+```powershell
+cd native-app
+dotnet run --project ./src/VoxType.NativeApp -c Release
+```
+
+首次启动点击“下载模型”。模型约 25 MB，保存到 `%LOCALAPPDATA%\VoxType\models\zipformer-zh-streaming-small`，不进入应用安装包。模型就绪后：
+
+1. 点击“开始实时识别”。
+2. 自然说话，结果框会持续显示临时结果；短暂停顿后自动确认一段。
+3. 点击“停止识别”提交最后一段，然后复制结果。
+
+麦克风每 40 ms 产生一个 PCM chunk，由后台在线识别会话持续消费，WPF UI 只接收 partial/final 更新，不在录音回调或 UI 线程执行模型推理。当前预览版不注册全局热键、不自动向其它窗口输入文字。
+
+## 构建小体积包
+
+```powershell
+pwsh ./build.ps1
+```
+
+产物：
+
+- `publish/win-x64/`：framework-dependent Windows x64 应用。
+- `publish/VoxType.NativePreview-win-x64.zip`：可直接解压体验的压缩包。
+
+实测 zip 为 8.1 MB。5 秒中文音频按 40 ms 实时分块播放时产生 8 次文字更新，首个 partial 约 0.91 秒出现；环境、内存与复现命令见 [`BENCHMARKS.md`](BENCHMARKS.md)。
+
+每次 `main` 上的 native-app 改动也会生成一个保留 14 天的 GitHub Actions artifact：打开仓库的 **Actions → Native Preview**，下载 `VoxType-NativePreview-win-x64`。
+
+应用要求已安装 .NET 10 Desktop Runtime。Quicker V2 已使用同一运行时；若未来面向没有 Quicker 的独立用户发布，可另做 self-contained 包，但体积会显著增加。
+
+## 命令行识别验证
+
+```powershell
+dotnet run --project ./src/VoxType.NativeSmoke -c Release -- <audio.wav> [model-directory] [repeat-count] [--realtime]
+```
+
+第二个可选参数是模型目录，第三个参数可在同一进程重复识别。第四个参数 `--realtime` 按真实 40 ms 节奏送入音频，用于检查 partial 次数和首字延迟。该工具用于 CI/开发验证，不进入预览应用压缩包。
+
+## 当前边界
+
+- 当前小模型只支持中文，没有 SenseVoice 的多语种、情绪识别和自动标点能力；换取约 89% 的权重体积下降与更低内存。
+- 只启用 Windows x64 CPU provider，优先保证部署一致性；识别器在进程内只加载一次，每次录音创建独立在线 stream。
+- 模型使用组件内独立 catalog，按文件大小与 SHA-256 校验，并支持 `.part` 断点续传。
+- 会话为首字补充 300 ms 左侧上下文；端点检测在连续静音后确认当前分段，并对重置边界的长文本重叠去重。点击停止时再送入 600 ms 右侧上下文刷新最终结果。
+- 模型来自 sherpa-onnx 的 [`streaming-zipformer-small-ctc-zh-int8-2025-04-01`](https://k2-fsa.github.io/sherpa/onnx/pretrained_models/online-ctc/zipformer-ctc-models.html)，镜像标记为 Apache-2.0；正式集成前仍需保留模型 attribution。

native-app/VoxType.Native.slnx

@@ -0,0 +1,5 @@
+<Solution>
+  <Project Path="src/VoxType.NativeCore/VoxType.NativeCore.csproj" />
+  <Project Path="src/VoxType.NativeApp/VoxType.NativeApp.csproj" />
+  <Project Path="src/VoxType.NativeSmoke/VoxType.NativeSmoke.csproj" />
+</Solution>

native-app/build.ps1

@@ -0,0 +1,50 @@
+#Requires -Version 7.0
+param(
+    [switch]$SkipBuild
+)
+
+$ErrorActionPreference = "Stop"
+$Root = $PSScriptRoot
+$Solution = Join-Path $Root "VoxType.Native.slnx"
+$AppProject = Join-Path $Root "src/VoxType.NativeApp/VoxType.NativeApp.csproj"
+$PublishDir = Join-Path $Root "publish/win-x64"
+$ZipPath = Join-Path $Root "publish/VoxType.NativePreview-win-x64.zip"
+
+$ResolvedRoot = [System.IO.Path]::GetFullPath($Root).TrimEnd([System.IO.Path]::DirectorySeparatorChar)
+$ResolvedPublishDir = [System.IO.Path]::GetFullPath($PublishDir)
+if (-not $ResolvedPublishDir.StartsWith($ResolvedRoot + [System.IO.Path]::DirectorySeparatorChar, [System.StringComparison]::OrdinalIgnoreCase)) {
+    throw "Publish directory escaped native-app root: $ResolvedPublishDir"
+}
+if (Test-Path -LiteralPath $ResolvedPublishDir) {
+    $PublishItem = Get-Item -LiteralPath $ResolvedPublishDir -Force
+    if ($PublishItem.Attributes -band [System.IO.FileAttributes]::ReparsePoint) {
+        throw "Refusing to clean reparse-point publish directory: $ResolvedPublishDir"
+    }
+    Remove-Item -LiteralPath $ResolvedPublishDir -Recurse -Force
+}
+
+if (-not $SkipBuild) {
+    dotnet build $Solution -c Release
+    if ($LASTEXITCODE -ne 0) { throw "Native app build failed." }
+}
+
+dotnet publish $AppProject `
+    -c Release `
+    -r win-x64 `
+    --self-contained false `
+    --no-restore `
+    -p:DebugType=None `
+    -p:DebugSymbols=false `
+    -o $ResolvedPublishDir
+if ($LASTEXITCODE -ne 0) { throw "Native app publish failed." }
+
+if (Test-Path -LiteralPath $ZipPath) {
+    Remove-Item -LiteralPath $ZipPath -Force
+}
+Compress-Archive -Path (Join-Path $PublishDir "*") -DestinationPath $ZipPath -CompressionLevel Optimal
+
+$PublishBytes = (Get-ChildItem -LiteralPath $PublishDir -Recurse -File | Measure-Object Length -Sum).Sum
+$ZipBytes = (Get-Item -LiteralPath $ZipPath).Length
+Write-Host "Native preview ready:" -ForegroundColor Green
+Write-Host "  Directory: $PublishDir ($([math]::Round($PublishBytes / 1MB, 1)) MB)"
+Write-Host "  Zip:       $ZipPath ($([math]::Round($ZipBytes / 1MB, 1)) MB)"

native-app/catalog/models.json

@@ -0,0 +1,34 @@
+{
+  "schemaVersion": 1,
+  "notes": "Native Preview streaming model catalog. Model weights are downloaded on first use and are not bundled with the app.",
+  "models": [
+    {
+      "id": "zipformer-streaming-small-zh-int8",
+      "name": "Zipformer 中文实时 Small",
+      "description": "推荐 · 约 25 MB · 中文流式识别",
+      "type": "streaming-zipformer-ctc",
+      "runtimePreset": "streaming-zipformer-ctc",
+      "layout": "zipformer-zh-streaming-small",
+      "supported": true,
+      "default": true,
+      "download": {
+        "source": "modelscope",
+        "modelscopeResolveBase": "https://www.modelscope.cn/models/manyeyes/k2transducer-zipformer-ctc-small-zh-onnx-online-20250401/resolve/master",
+        "modelscopeFiles": [
+          {
+            "name": "model.int8.onnx",
+            "size": 26342340,
+            "required": true,
+            "sha256": "68c9c943840f7d9cf3e8a4970ba50f404feb5277f611fa82b7e72267786fa84a"
+          },
+          {
+            "name": "tokens.txt",
+            "size": 13366,
+            "required": true,
+            "sha256": "6fed8c6c248516f38e7faa19404b57413e8ce259f1cbc1fa4aebc86eac32fdfd"
+          }
+        ]
+      }
+    }
+  ]
+}

native-app/src/VoxType.NativeApp/App.xaml

@@ -0,0 +1,50 @@
+<Application x:Class="VoxType.NativeApp.App"
+             xmlns="http://schemas.microsoft.com/winfx/2006/xaml/presentation"
+             xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml"
+             StartupUri="MainWindow.xaml">
+    <Application.Resources>
+        <SolidColorBrush x:Key="WindowBackgroundBrush" Color="#0B0E14" />
+        <SolidColorBrush x:Key="PanelBrush" Color="#141923" />
+        <SolidColorBrush x:Key="PanelAltBrush" Color="#1A2130" />
+        <SolidColorBrush x:Key="BorderBrush" Color="#2A354A" />
+        <SolidColorBrush x:Key="TextBrush" Color="#EEF2F7" />
+        <SolidColorBrush x:Key="MutedTextBrush" Color="#98A4B7" />
+        <SolidColorBrush x:Key="AccentBrush" Color="#4F8CFF" />
+        <SolidColorBrush x:Key="AccentHoverBrush" Color="#6BA0FF" />
+
+        <Style TargetType="TextBlock">
+            <Setter Property="Foreground" Value="{StaticResource TextBrush}" />
+        </Style>
+        <Style TargetType="Button">
+            <Setter Property="Foreground" Value="White" />
+            <Setter Property="Background" Value="{StaticResource AccentBrush}" />
+            <Setter Property="BorderThickness" Value="0" />
+            <Setter Property="Padding" Value="16,9" />
+            <Setter Property="FontWeight" Value="SemiBold" />
+            <Setter Property="Cursor" Value="Hand" />
+            <Setter Property="Template">
+                <Setter.Value>
+                    <ControlTemplate TargetType="Button">
+                        <Border x:Name="Root"
+                                Padding="{TemplateBinding Padding}"
+                                Background="{TemplateBinding Background}"
+                                CornerRadius="9">
+                            <ContentPresenter HorizontalAlignment="Center" VerticalAlignment="Center" />
+                        </Border>
+                        <ControlTemplate.Triggers>
+                            <Trigger Property="IsMouseOver" Value="True">
+                                <Setter TargetName="Root" Property="Background" Value="{StaticResource AccentHoverBrush}" />
+                            </Trigger>
+                            <Trigger Property="IsEnabled" Value="False">
+                                <Setter TargetName="Root" Property="Opacity" Value="0.45" />
+                            </Trigger>
+                        </ControlTemplate.Triggers>
+                    </ControlTemplate>
+                </Setter.Value>
+            </Setter>
+        </Style>
+        <Style x:Key="SecondaryButton" TargetType="Button" BasedOn="{StaticResource {x:Type Button}}">
+            <Setter Property="Background" Value="#273247" />
+        </Style>
+    </Application.Resources>
+</Application>

native-app/src/VoxType.NativeApp/App.xaml.cs

@@ -0,0 +1,7 @@
+using System.Windows;
+
+namespace VoxType.NativeApp;
+
+public partial class App : Application
+{
+}