From 2bc530679549fc17af4ddef79f31160b45feb47f Mon Sep 17 00:00:00 2001
From: Cea <61349137+ceastld@users.noreply.github.com>
Date: Fri, 7 Aug 2026 19:47:50 +0800
Subject: [PATCH 1/4] feat: add lightweight native voice preview

---
 .github/workflows/native-preview.yml          |  46 +++
 .gitignore                                    |   5 +
 AGENTS.md                                     |   6 +
 README.md                                     |   3 +
 docs/quicker-integration-research.md          |   2 +
 native-app/BENCHMARKS.md                      |  43 +++
 native-app/Directory.Build.props              |  10 +
 native-app/Directory.Build.targets            |   3 +
 native-app/Directory.Packages.props           |  10 +
 native-app/README.md                          |  50 ++++
 native-app/VoxType.Native.slnx                |   5 +
 native-app/build.ps1                          |  50 ++++
 native-app/src/VoxType.NativeApp/App.xaml     |  50 ++++
 native-app/src/VoxType.NativeApp/App.xaml.cs  |   7 +
 .../src/VoxType.NativeApp/MainWindow.xaml     | 172 +++++++++++
 .../src/VoxType.NativeApp/MainWindow.xaml.cs  | 278 ++++++++++++++++++
 .../VoxType.NativeApp.csproj                  |  22 ++
 .../VoxType.NativeCore/AudioCaptureSession.cs | 110 +++++++
 .../src/VoxType.NativeCore/ModelCatalog.cs    |  91 ++++++
 .../src/VoxType.NativeCore/ModelManager.cs    | 206 +++++++++++++
 .../VoxType.NativeCore/SenseVoiceEngine.cs    | 124 ++++++++
 .../VoxType.NativeCore.csproj                 |  14 +
 .../src/VoxType.NativeCore/VoxTypePaths.cs    |  37 +++
 .../src/VoxType.NativeCore/WaveFileAudio.cs   |  42 +++
 native-app/src/VoxType.NativeSmoke/Program.cs |  56 ++++
 .../VoxType.NativeSmoke.csproj                |  14 +
 26 files changed, 1456 insertions(+)
 create mode 100644 .github/workflows/native-preview.yml
 create mode 100644 native-app/BENCHMARKS.md
 create mode 100644 native-app/Directory.Build.props
 create mode 100644 native-app/Directory.Build.targets
 create mode 100644 native-app/Directory.Packages.props
 create mode 100644 native-app/README.md
 create mode 100644 native-app/VoxType.Native.slnx
 create mode 100644 native-app/build.ps1
 create mode 100644 native-app/src/VoxType.NativeApp/App.xaml
 create mode 100644 native-app/src/VoxType.NativeApp/App.xaml.cs
 create mode 100644 native-app/src/VoxType.NativeApp/MainWindow.xaml
 create mode 100644 native-app/src/VoxType.NativeApp/MainWindow.xaml.cs
 create mode 100644 native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj
 create mode 100644 native-app/src/VoxType.NativeCore/AudioCaptureSession.cs
 create mode 100644 native-app/src/VoxType.NativeCore/ModelCatalog.cs
 create mode 100644 native-app/src/VoxType.NativeCore/ModelManager.cs
 create mode 100644 native-app/src/VoxType.NativeCore/SenseVoiceEngine.cs
 create mode 100644 native-app/src/VoxType.NativeCore/VoxType.NativeCore.csproj
 create mode 100644 native-app/src/VoxType.NativeCore/VoxTypePaths.cs
 create mode 100644 native-app/src/VoxType.NativeCore/WaveFileAudio.cs
 create mode 100644 native-app/src/VoxType.NativeSmoke/Program.cs
 create mode 100644 native-app/src/VoxType.NativeSmoke/VoxType.NativeSmoke.csproj

diff --git a/.github/workflows/native-preview.yml b/.github/workflows/native-preview.yml
new file mode 100644
index 0000000..3ec1d80
--- /dev/null
+++ b/.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@v4
+
+      - name: Setup .NET 10
+        uses: actions/setup-dotnet@v4
+        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@v4
+        with:
+          name: VoxType-NativePreview-win-x64
+          path: native-app/publish/VoxType.NativePreview-win-x64.zip
+          if-no-files-found: error
+          retention-days: 14
diff --git a/.gitignore b/.gitignore
index b943aa1..dc24b8c 100644
--- a/.gitignore
+++ b/.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/
diff --git a/AGENTS.md b/AGENTS.md
index 49f3bf0..d5066db 100644
--- a/AGENTS.md
+++ b/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
diff --git a/README.md b/README.md
index 7de0e59..6bcffcb 100644
--- a/README.md
+++ b/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.
diff --git a/docs/quicker-integration-research.md b/docs/quicker-integration-research.md
index 5b37db1..77a6369 100644
--- a/docs/quicker-integration-research.md
+++ b/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，用于验证更小的原生 worker；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
 
diff --git a/native-app/BENCHMARKS.md b/native-app/BENCHMARKS.md
new file mode 100644
index 0000000..50ffcaf
--- /dev/null
+++ b/native-app/BENCHMARKS.md
@@ -0,0 +1,43 @@
+# 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，CPU provider，最多 4 线程
+- 模型：SenseVoice Small INT8，239,233,841-byte ONNX
+- 音频：16 kHz mono，5.00 秒中文 WAV
+
+## 结果
+
+| 指标 | 结果 |
+|------|------|
+| 模型加载 | 1.27–1.63 s |
+| 稳态识别 | 0.16–0.23 s，RTF 0.032–0.045 |
+| 发布目录 | 21.4 MB |
+| zip | 8.1 MB |
+| 模型加载后的进程 Working Set | 约 506 MB |
+| 模型加载后的 Private Memory | 约 450 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。重复运行存在正常调度波动，因此表中保留两轮实测范围。
+
+## 解释
+
+- 首次加载和后续识别必须分开测量；UI 启动时预热模型，因此用户松开录音后的主要等待接近“稳态识别”。
+- 体积下降主要来自移除 Python、NumPy、aiohttp 和 PyInstaller runtime；发布目录仍主要由 `onnxruntime.dll` 与 `sherpa-onnx-c-api.dll` 构成。
+- 内存主要由 229 MiB 模型、ONNX Runtime 权重/工作区和 WPF 进程构成。后续 Quicker 接入仍建议使用独立原生 worker，以隔离这部分内存和 native DLL 生命周期。
+- 当前只启用 CPU provider。DirectML/CUDA 是否值得加入，要以更多硬件上的启动失败率、驱动兼容和安装包增量为准。
+
+## 复现
+
+```powershell
+pwsh ./build.ps1
+
+dotnet run --project ./src/VoxType.NativeSmoke -c Release -- `
+  <audio.wav> `
+  "$env:LOCALAPPDATA\VoxType\models\sensevoice" `
+  3
+```
diff --git a/native-app/Directory.Build.props b/native-app/Directory.Build.props
new file mode 100644
index 0000000..b377359
--- /dev/null
+++ b/native-app/Directory.Build.props
@@ -0,0 +1,10 @@
+<Project>
+  <!-- Keep this standalone project isolated from Directory.Build.props files in parent checkouts. -->
+  <PropertyGroup>
+    <LangVersion>latest</LangVersion>
+    <TreatWarningsAsErrors>true</TreatWarningsAsErrors>
+    <DefaultItemExcludes>$(DefaultItemExcludes);publish/**;**/publish/**</DefaultItemExcludes>
+    <QkDisableQkref>true</QkDisableQkref>
+    <QkDisablePolySharp>true</QkDisablePolySharp>
+  </PropertyGroup>
+</Project>
diff --git a/native-app/Directory.Build.targets b/native-app/Directory.Build.targets
new file mode 100644
index 0000000..a25c1a7
--- /dev/null
+++ b/native-app/Directory.Build.targets
@@ -0,0 +1,3 @@
+<Project>
+  <!-- Intentionally empty: stop parent repositories from injecting build targets. -->
+</Project>
diff --git a/native-app/Directory.Packages.props b/native-app/Directory.Packages.props
new file mode 100644
index 0000000..d4515eb
--- /dev/null
+++ b/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>
diff --git a/native-app/README.md b/native-app/README.md
new file mode 100644
index 0000000..87bafe8
--- /dev/null
+++ b/native-app/README.md
@@ -0,0 +1,50 @@
+# VoxType Native Preview
+
+面向 `QuickerPc/Quicker` 的最小原生语音识别体验版。它使用 `.NET 10 + WPF + NAudio + sherpa-onnx C#`，直接加载 SenseVoice，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 229 MB，保存到 `%LOCALAPPDATA%\VoxType\models\sensevoice`，与现有 VoxType 共用，不进入应用安装包。模型就绪后：
+
+1. 点击“开始录音”。
+2. 说话后点击“结束并识别”。
+3. 查看耗时和 RTF，复制识别结果。
+
+当前预览版只做录音、识别和复制，不注册全局热键、不自动向其它窗口输入文字。这让首轮测试集中在原生运行稳定性、识别性能和打包体积。
+
+## 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。
+
+首轮实测 zip 为 8.1 MB，模型加载后的 5 秒中文音频稳态识别为 0.16–0.23 秒。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]
+```
+
+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，用于区分模型加载耗时和稳态 decode 耗时。该工具用于 CI/开发验证，不进入预览应用压缩包。
+
+## 当前边界
+
+- 只启用 Windows x64 CPU provider，优先保证部署一致性；GPU provider 留待实测收益与驱动兼容性后再开放。
+- 识别器在进程内只加载一次，每次录音创建独立 stream，避免重复加载模型。
+- 模型按 catalog 下载并支持 `.part` 断点续传。
+- 原生识别调用不可被 sherpa-onnx 中途打断；关闭窗口会阻止新会话，但正在执行的单次 decode 需要自然返回。
diff --git a/native-app/VoxType.Native.slnx b/native-app/VoxType.Native.slnx
new file mode 100644
index 0000000..91669aa
--- /dev/null
+++ b/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>
diff --git a/native-app/build.ps1 b/native-app/build.ps1
new file mode 100644
index 0000000..9e63889
--- /dev/null
+++ b/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)"
diff --git a/native-app/src/VoxType.NativeApp/App.xaml b/native-app/src/VoxType.NativeApp/App.xaml
new file mode 100644
index 0000000..88fbe7f
--- /dev/null
+++ b/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>
diff --git a/native-app/src/VoxType.NativeApp/App.xaml.cs b/native-app/src/VoxType.NativeApp/App.xaml.cs
new file mode 100644
index 0000000..e10c9d8
--- /dev/null
+++ b/native-app/src/VoxType.NativeApp/App.xaml.cs
@@ -0,0 +1,7 @@
+using System.Windows;
+
+namespace VoxType.NativeApp;
+
+public partial class App : Application
+{
+}
diff --git a/native-app/src/VoxType.NativeApp/MainWindow.xaml b/native-app/src/VoxType.NativeApp/MainWindow.xaml
new file mode 100644
index 0000000..d4d7142
--- /dev/null
+++ b/native-app/src/VoxType.NativeApp/MainWindow.xaml
@@ -0,0 +1,172 @@
+<Window x:Class="VoxType.NativeApp.MainWindow"
+        xmlns="http://schemas.microsoft.com/winfx/2006/xaml/presentation"
+        xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml"
+        Title="VoxType Native Preview"
+        Width="760"
+        Height="650"
+        MinWidth="660"
+        MinHeight="560"
+        WindowStartupLocation="CenterScreen"
+        Background="{StaticResource WindowBackgroundBrush}"
+        Foreground="{StaticResource TextBrush}"
+        FontFamily="Segoe UI"
+        Loaded="Window_Loaded"
+        Closing="Window_Closing">
+    <Grid Margin="24">
+        <Grid.RowDefinitions>
+            <RowDefinition Height="Auto" />
+            <RowDefinition Height="Auto" />
+            <RowDefinition Height="Auto" />
+            <RowDefinition Height="*" />
+            <RowDefinition Height="Auto" />
+        </Grid.RowDefinitions>
+
+        <Grid>
+            <Grid.ColumnDefinitions>
+                <ColumnDefinition Width="*" />
+                <ColumnDefinition Width="Auto" />
+            </Grid.ColumnDefinitions>
+            <StackPanel>
+                <TextBlock FontSize="26" FontWeight="SemiBold" Text="VoxType Native" />
+                <TextBlock Margin="0,4,0,0"
+                           Foreground="{StaticResource MutedTextBrush}"
+                           Text="无 Python · 原生 sherpa-onnx · 面向 Quicker 的最小体验版" />
+            </StackPanel>
+            <Border Grid.Column="1"
+                    Padding="12,7"
+                    VerticalAlignment="Center"
+                    Background="#173A2A"
+                    CornerRadius="99">
+                <TextBlock Foreground="#79E29E" FontWeight="SemiBold" Text="CPU · 本地离线" />
+            </Border>
+        </Grid>
+
+        <Border Grid.Row="1"
+                Margin="0,22,0,0"
+                Padding="16"
+                Background="{StaticResource PanelBrush}"
+                BorderBrush="{StaticResource BorderBrush}"
+                BorderThickness="1"
+                CornerRadius="14">
+            <Grid>
+                <Grid.ColumnDefinitions>
+                    <ColumnDefinition Width="*" />
+                    <ColumnDefinition Width="Auto" />
+                </Grid.ColumnDefinitions>
+                <StackPanel>
+                    <TextBlock FontSize="15" FontWeight="SemiBold" Text="SenseVoice Small INT8" />
+                    <TextBlock x:Name="ModelStatusText"
+                               Margin="0,5,0,0"
+                               Foreground="{StaticResource MutedTextBrush}"
+                               Text="正在检查模型…" />
+                    <ProgressBar x:Name="DownloadProgressBar"
+                                 Height="5"
+                                 Margin="0,12,0,0"
+                                 Maximum="100"
+                                 Visibility="Collapsed" />
+                </StackPanel>
+                <StackPanel Grid.Column="1" Margin="18,0,0,0" Orientation="Horizontal" VerticalAlignment="Center">
+                    <Button x:Name="OpenModelFolderButton"
+                            Margin="0,0,8,0"
+                            Click="OpenModelFolderButton_Click"
+                            Style="{StaticResource SecondaryButton}"
+                            Content="打开目录" />
+                    <Button x:Name="DownloadModelButton"
+                            Click="DownloadModelButton_Click"
+                            Content="下载模型" />
+                </StackPanel>
+            </Grid>
+        </Border>
+
+        <Border Grid.Row="2"
+                Margin="0,14,0,0"
+                Padding="18"
+                Background="{StaticResource PanelBrush}"
+                BorderBrush="{StaticResource BorderBrush}"
+                BorderThickness="1"
+                CornerRadius="14">
+            <Grid>
+                <Grid.ColumnDefinitions>
+                    <ColumnDefinition Width="Auto" />
+                    <ColumnDefinition Width="*" />
+                    <ColumnDefinition Width="Auto" />
+                </Grid.ColumnDefinitions>
+                <Button x:Name="RecordButton"
+                        Width="150"
+                        Height="54"
+                        IsEnabled="False"
+                        Click="RecordButton_Click"
+                        Content="开始录音" />
+                <StackPanel Grid.Column="1" Margin="18,0" VerticalAlignment="Center">
+                    <TextBlock x:Name="SessionTitleText" FontSize="15" FontWeight="SemiBold" Text="准备就绪" />
+                    <TextBlock x:Name="SessionHintText"
+                               Margin="0,4,0,0"
+                               Foreground="{StaticResource MutedTextBrush}"
+                               Text="点击开始录音，再次点击结束并识别" />
+                </StackPanel>
+                <TextBlock x:Name="DurationText"
+                           Grid.Column="2"
+                           VerticalAlignment="Center"
+                           FontFamily="Consolas"
+                           FontSize="22"
+                           Foreground="#8AB4FF"
+                           Text="00:00.0" />
+            </Grid>
+        </Border>
+
+        <Border Grid.Row="3"
+                Margin="0,14,0,0"
+                Padding="16"
+                Background="{StaticResource PanelBrush}"
+                BorderBrush="{StaticResource BorderBrush}"
+                BorderThickness="1"
+                CornerRadius="14">
+            <Grid>
+                <Grid.RowDefinitions>
+                    <RowDefinition Height="Auto" />
+                    <RowDefinition Height="*" />
+                    <RowDefinition Height="Auto" />
+                </Grid.RowDefinitions>
+                <Grid>
+                    <TextBlock FontSize="14" FontWeight="SemiBold" Text="识别结果" />
+                    <StackPanel HorizontalAlignment="Right" Orientation="Horizontal">
+                        <Button x:Name="ClearButton"
+                                Margin="0,0,8,0"
+                                Padding="12,6"
+                                Click="ClearButton_Click"
+                                Style="{StaticResource SecondaryButton}"
+                                Content="清空" />
+                        <Button x:Name="CopyButton"
+                                Padding="12,6"
+                                IsEnabled="False"
+                                Click="CopyButton_Click"
+                                Content="复制" />
+                    </StackPanel>
+                </Grid>
+                <TextBox x:Name="ResultTextBox"
+                         Grid.Row="1"
+                         Margin="0,14,0,0"
+                         Padding="14"
+                         AcceptsReturn="True"
+                         Background="{StaticResource PanelAltBrush}"
+                         BorderBrush="{StaticResource BorderBrush}"
+                         Foreground="{StaticResource TextBrush}"
+                         FontSize="18"
+                         IsReadOnly="True"
+                         TextWrapping="Wrap"
+                         VerticalScrollBarVisibility="Auto" />
+                <TextBlock x:Name="MetricsText"
+                           Grid.Row="2"
+                           Margin="0,10,0,0"
+                           Foreground="{StaticResource MutedTextBrush}"
+                           Text="模型和音频均只保存在本机" />
+            </Grid>
+        </Border>
+
+        <TextBlock x:Name="StatusText"
+                   Grid.Row="4"
+                   Margin="2,14,0,0"
+                   Foreground="{StaticResource MutedTextBrush}"
+                   Text="初始化…" />
+    </Grid>
+</Window>
diff --git a/native-app/src/VoxType.NativeApp/MainWindow.xaml.cs b/native-app/src/VoxType.NativeApp/MainWindow.xaml.cs
new file mode 100644
index 0000000..3df69d5
--- /dev/null
+++ b/native-app/src/VoxType.NativeApp/MainWindow.xaml.cs
@@ -0,0 +1,278 @@
+using System.ComponentModel;
+using System.Diagnostics;
+using System.IO;
+using System.Windows;
+using System.Windows.Media;
+using System.Windows.Threading;
+using VoxType.NativeCore;
+
+namespace VoxType.NativeApp;
+
+public partial class MainWindow : Window
+{
+    private readonly SenseVoiceEngine _engine = new();
+    private readonly ModelManager _modelManager = new();
+    private readonly CancellationTokenSource _lifetime = new();
+    private readonly DispatcherTimer _recordingTimer;
+    private ModelCatalogEntry? _model;
+    private AudioCaptureSession? _capture;
+    private Stopwatch? _recordingWatch;
+    private bool _busy;
+
+    public MainWindow()
+    {
+        InitializeComponent();
+        _recordingTimer = new DispatcherTimer(TimeSpan.FromMilliseconds(100), DispatcherPriority.Background, (_, _) =>
+        {
+            var elapsed = _recordingWatch?.Elapsed ?? TimeSpan.Zero;
+            DurationText.Text = $"{(int)elapsed.TotalMinutes:00}:{elapsed.Seconds:00}.{elapsed.Milliseconds / 100}";
+        }, Dispatcher);
+    }
+
+    private async void Window_Loaded(object sender, RoutedEventArgs e)
+    {
+        try
+        {
+            var catalogPath = Path.Combine(AppContext.BaseDirectory, "catalog", "models.json");
+            _model = ModelCatalog.Load(catalogPath).GetDefaultSenseVoice();
+            RefreshModelState();
+            if (ModelManager.IsInstalled(_model))
+            {
+                await WarmUpAsync();
+            }
+            else
+            {
+                StatusText.Text = "首次使用需下载约 229 MB 的模型文件，模型不会打进应用包。";
+            }
+        }
+        catch (Exception ex)
+        {
+            ShowError("初始化失败", ex);
+        }
+    }
+
+    private async void DownloadModelButton_Click(object sender, RoutedEventArgs e)
+    {
+        if (_model is null || _busy)
+        {
+            return;
+        }
+
+        SetBusy(true);
+        DownloadProgressBar.Visibility = Visibility.Visible;
+        DownloadProgressBar.Value = 0;
+        try
+        {
+            var progress = new Progress<ModelDownloadProgress>(value =>
+            {
+                DownloadProgressBar.Value = value.Percentage;
+                ModelStatusText.Text = value.Message;
+                StatusText.Text = value.Message;
+            });
+            await _modelManager.EnsureInstalledAsync(_model, progress, _lifetime.Token);
+            RefreshModelState();
+            await WarmUpAsync();
+        }
+        catch (OperationCanceledException)
+        {
+            StatusText.Text = "下载已取消，可再次点击继续。";
+        }
+        catch (Exception ex)
+        {
+            ShowError("模型下载失败", ex);
+        }
+        finally
+        {
+            DownloadProgressBar.Visibility = Visibility.Collapsed;
+            SetBusy(false);
+        }
+    }
+
+    private async Task WarmUpAsync()
+    {
+        if (_model is null)
+        {
+            return;
+        }
+
+        SetBusy(true);
+        SessionTitleText.Text = "正在加载模型";
+        SessionHintText.Text = "首次加载通常需要几秒";
+        StatusText.Text = "正在创建原生 SenseVoice 识别器…";
+        var watch = Stopwatch.StartNew();
+        try
+        {
+            await _engine.WarmUpAsync(VoxTypePaths.ResolveModelDirectory(_model), _lifetime.Token);
+            SessionTitleText.Text = "准备就绪";
+            SessionHintText.Text = "点击开始录音，再次点击结束并识别";
+            StatusText.Text = $"模型已加载 · {watch.Elapsed.TotalSeconds:F2}s";
+        }
+        finally
+        {
+            SetBusy(false);
+        }
+    }
+
+    private async void RecordButton_Click(object sender, RoutedEventArgs e)
+    {
+        if (_busy || _model is null || !ModelManager.IsInstalled(_model))
+        {
+            return;
+        }
+
+        if (_capture is null)
+        {
+            StartRecording();
+        }
+        else
+        {
+            await StopAndTranscribeAsync();
+        }
+    }
+
+    private void StartRecording()
+    {
+        try
+        {
+            _capture = AudioCaptureSession.Start();
+            _recordingWatch = Stopwatch.StartNew();
+            _recordingTimer.Start();
+            RecordButton.Content = "结束并识别";
+            RecordButton.Background = new SolidColorBrush(Color.FromRgb(213, 75, 75));
+            SessionTitleText.Text = "正在录音";
+            SessionHintText.Text = "自然说话，结束后会在本机识别";
+            StatusText.Text = "麦克风录音中…";
+        }
+        catch (Exception ex)
+        {
+            ShowError("无法开始录音", ex);
+        }
+    }
+
+    private async Task StopAndTranscribeAsync()
+    {
+        var capture = _capture;
+        _capture = null;
+        if (capture is null || _model is null)
+        {
+            return;
+        }
+
+        _recordingTimer.Stop();
+        _recordingWatch?.Stop();
+        SetBusy(true);
+        RecordButton.Content = "开始录音";
+        RecordButton.ClearValue(BackgroundProperty);
+        SessionTitleText.Text = "正在识别";
+        SessionHintText.Text = "原生推理不会上传音频";
+        StatusText.Text = "正在运行 SenseVoice…";
+
+        try
+        {
+            var samples = await capture.StopAsync(_lifetime.Token);
+            capture.Dispose();
+            var audioSeconds = samples.Length / (double)AudioCaptureSession.SampleRate;
+            if (audioSeconds < 0.25)
+            {
+                throw new InvalidOperationException("录音太短，请至少说 0.3 秒。");
+            }
+
+            var recognitionWatch = Stopwatch.StartNew();
+            var text = await _engine.TranscribeAsync(
+                VoxTypePaths.ResolveModelDirectory(_model),
+                samples,
+                _lifetime.Token);
+            recognitionWatch.Stop();
+
+            ResultTextBox.Text = text;
+            CopyButton.IsEnabled = !string.IsNullOrWhiteSpace(text);
+            MetricsText.Text = $"音频 {audioSeconds:F1}s · 识别 {recognitionWatch.Elapsed.TotalSeconds:F2}s · RTF {recognitionWatch.Elapsed.TotalSeconds / audioSeconds:F2}";
+            SessionTitleText.Text = string.IsNullOrWhiteSpace(text) ? "未识别到文字" : "识别完成";
+            SessionHintText.Text = "可复制结果，或继续下一段录音";
+            StatusText.Text = string.IsNullOrWhiteSpace(text) ? "没有识别结果，请靠近麦克风后重试。" : "识别结果已生成。";
+        }
+        catch (OperationCanceledException)
+        {
+            StatusText.Text = "本次识别已取消。";
+        }
+        catch (Exception ex)
+        {
+            ShowError("识别失败", ex);
+        }
+        finally
+        {
+            capture.Dispose();
+            SetBusy(false);
+        }
+    }
+
+    private void CopyButton_Click(object sender, RoutedEventArgs e)
+    {
+        if (string.IsNullOrWhiteSpace(ResultTextBox.Text))
+        {
+            return;
+        }
+
+        Clipboard.SetText(ResultTextBox.Text);
+        StatusText.Text = "识别结果已复制。";
+    }
+
+    private void ClearButton_Click(object sender, RoutedEventArgs e)
+    {
+        ResultTextBox.Clear();
+        CopyButton.IsEnabled = false;
+        MetricsText.Text = "模型和音频均只保存在本机";
+    }
+
+    private void OpenModelFolderButton_Click(object sender, RoutedEventArgs e)
+    {
+        Directory.CreateDirectory(VoxTypePaths.ModelsRoot);
+        Process.Start(new ProcessStartInfo
+        {
+            FileName = "explorer.exe",
+            ArgumentList = { VoxTypePaths.ModelsRoot },
+            UseShellExecute = true,
+        });
+    }
+
+    private void RefreshModelState()
+    {
+        if (_model is null)
+        {
+            return;
+        }
+
+        var installed = ModelManager.IsInstalled(_model);
+        ModelStatusText.Text = installed
+            ? $"已安装 · {VoxTypePaths.ResolveModelDirectory(_model)}"
+            : "未安装 · 首次使用下载约 229 MB";
+        DownloadModelButton.Content = installed ? "校验/续传" : "下载模型";
+        RecordButton.IsEnabled = installed && !_busy;
+    }
+
+    private void SetBusy(bool busy)
+    {
+        _busy = busy;
+        DownloadModelButton.IsEnabled = !busy;
+        OpenModelFolderButton.IsEnabled = !busy;
+        RecordButton.IsEnabled = !busy && _model is not null && ModelManager.IsInstalled(_model);
+    }
+
+    private void ShowError(string title, Exception ex)
+    {
+        SessionTitleText.Text = title;
+        SessionHintText.Text = ex.Message;
+        StatusText.Text = $"{title}：{ex.Message}";
+    }
+
+    private void Window_Closing(object? sender, CancelEventArgs e)
+    {
+        _lifetime.Cancel();
+        _recordingTimer.Stop();
+        _capture?.Dispose();
+        _capture = null;
+        _engine.Dispose();
+        _modelManager.Dispose();
+        _lifetime.Dispose();
+    }
+}
diff --git a/native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj b/native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj
new file mode 100644
index 0000000..19af92d
--- /dev/null
+++ b/native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj
@@ -0,0 +1,22 @@
+<Project Sdk="Microsoft.NET.Sdk">
+  <PropertyGroup>
+    <OutputType>WinExe</OutputType>
+    <TargetFramework>net10.0-windows</TargetFramework>
+    <UseWPF>true</UseWPF>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <Nullable>enable</Nullable>
+    <PlatformTarget>x64</PlatformTarget>
+    <RuntimeIdentifier>win-x64</RuntimeIdentifier>
+    <SelfContained>false</SelfContained>
+    <AssemblyName>VoxType.NativePreview</AssemblyName>
+    <ApplicationIcon>../../../app/src-tauri/icons/icon.ico</ApplicationIcon>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <ProjectReference Include="../VoxType.NativeCore/VoxType.NativeCore.csproj" />
+    <Content Include="../../../catalog/models.json" Link="catalog/models.json">
+      <CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
+      <CopyToPublishDirectory>PreserveNewest</CopyToPublishDirectory>
+    </Content>
+  </ItemGroup>
+</Project>
diff --git a/native-app/src/VoxType.NativeCore/AudioCaptureSession.cs b/native-app/src/VoxType.NativeCore/AudioCaptureSession.cs
new file mode 100644
index 0000000..547d78c
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/AudioCaptureSession.cs
@@ -0,0 +1,110 @@
+using NAudio.Wave;
+
+namespace VoxType.NativeCore;
+
+public sealed class AudioCaptureSession : IDisposable
+{
+    public const int SampleRate = 16_000;
+
+    private readonly object _gate = new();
+    private readonly List<float> _samples = [];
+    private readonly WaveInEvent _waveIn;
+    private readonly TaskCompletionSource _stopped = new(TaskCreationOptions.RunContinuationsAsynchronously);
+    private bool _isStopped;
+
+    private AudioCaptureSession(WaveInEvent waveIn)
+    {
+        _waveIn = waveIn;
+        _waveIn.DataAvailable += OnDataAvailable;
+        _waveIn.RecordingStopped += OnRecordingStopped;
+    }
+
+    public static AudioCaptureSession Start()
+    {
+        var waveIn = new WaveInEvent
+        {
+            DeviceNumber = 0,
+            WaveFormat = new WaveFormat(SampleRate, 16, 1),
+            BufferMilliseconds = 40,
+            NumberOfBuffers = 3,
+        };
+
+        var session = new AudioCaptureSession(waveIn);
+        try
+        {
+            waveIn.StartRecording();
+            return session;
+        }
+        catch
+        {
+            session.Dispose();
+            throw;
+        }
+    }
+
+    public async Task<float[]> StopAsync(CancellationToken cancellationToken = default)
+    {
+        lock (_gate)
+        {
+            if (!_isStopped)
+            {
+                _isStopped = true;
+                _waveIn.StopRecording();
+            }
+        }
+
+        await _stopped.Task.WaitAsync(cancellationToken).ConfigureAwait(false);
+        lock (_gate)
+        {
+            return [.. _samples];
+        }
+    }
+
+    private void OnDataAvailable(object? sender, WaveInEventArgs e)
+    {
+        lock (_gate)
+        {
+            for (var index = 0; index + 1 < e.BytesRecorded; index += 2)
+            {
+                var sample = (short)(e.Buffer[index] | e.Buffer[index + 1] << 8);
+                _samples.Add(sample / 32768f);
+            }
+        }
+    }
+
+    private void OnRecordingStopped(object? sender, StoppedEventArgs e)
+    {
+        if (e.Exception is not null)
+        {
+            _stopped.TrySetException(e.Exception);
+        }
+        else
+        {
+            _stopped.TrySetResult();
+        }
+    }
+
+    public void Dispose()
+    {
+        lock (_gate)
+        {
+            if (!_isStopped)
+            {
+                _isStopped = true;
+                try
+                {
+                    _waveIn.StopRecording();
+                }
+                catch
+                {
+                    // Best-effort shutdown during window close.
+                }
+            }
+        }
+
+        _waveIn.DataAvailable -= OnDataAvailable;
+        _waveIn.RecordingStopped -= OnRecordingStopped;
+        _waveIn.Dispose();
+        _stopped.TrySetResult();
+    }
+}
diff --git a/native-app/src/VoxType.NativeCore/ModelCatalog.cs b/native-app/src/VoxType.NativeCore/ModelCatalog.cs
new file mode 100644
index 0000000..9275f08
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/ModelCatalog.cs
@@ -0,0 +1,91 @@
+using System.Text.Json;
+using System.Text.Json.Serialization;
+
+namespace VoxType.NativeCore;
+
+public sealed class ModelCatalog
+{
+    [JsonPropertyName("models")]
+    public List<ModelCatalogEntry> Models { get; init; } = [];
+
+    public static ModelCatalog Load(string path)
+    {
+        if (!File.Exists(path))
+        {
+            throw new FileNotFoundException("未找到模型目录。", path);
+        }
+
+        var catalog = JsonSerializer.Deserialize<ModelCatalog>(File.ReadAllText(path), JsonOptions)
+            ?? throw new InvalidDataException("模型目录为空。");
+
+        if (catalog.Models.Count == 0)
+        {
+            throw new InvalidDataException("模型目录没有可用条目。");
+        }
+
+        return catalog;
+    }
+
+    public ModelCatalogEntry GetDefaultSenseVoice()
+    {
+        return Models.FirstOrDefault(model =>
+                   model.Supported && model.Default && model.Type.Equals("sensevoice", StringComparison.OrdinalIgnoreCase))
+               ?? Models.FirstOrDefault(model =>
+                   model.Supported && model.Type.Equals("sensevoice", StringComparison.OrdinalIgnoreCase))
+               ?? throw new InvalidDataException("模型目录未配置可用的 SenseVoice 模型。");
+    }
+
+    private static readonly JsonSerializerOptions JsonOptions = new()
+    {
+        PropertyNameCaseInsensitive = true,
+        ReadCommentHandling = JsonCommentHandling.Skip,
+    };
+}
+
+public sealed class ModelCatalogEntry
+{
+    [JsonPropertyName("id")]
+    public string Id { get; init; } = "";
+
+    [JsonPropertyName("name")]
+    public string Name { get; init; } = "";
+
+    [JsonPropertyName("description")]
+    public string Description { get; init; } = "";
+
+    [JsonPropertyName("type")]
+    public string Type { get; init; } = "";
+
+    [JsonPropertyName("layout")]
+    public string Layout { get; init; } = "";
+
+    [JsonPropertyName("supported")]
+    public bool Supported { get; init; }
+
+    [JsonPropertyName("default")]
+    public bool Default { get; init; }
+
+    [JsonPropertyName("download")]
+    public ModelDownloadSpec Download { get; init; } = new();
+}
+
+public sealed class ModelDownloadSpec
+{
+    [JsonPropertyName("modelscopeResolveBase")]
+    public string ModelScopeResolveBase { get; init; } = "";
+
+    [JsonPropertyName("modelscopeFiles")]
+    public List<ModelFileSpec> ModelScopeFiles { get; init; } = [];
+}
+
+public sealed class ModelFileSpec
+{
+    [JsonPropertyName("name")]
+    public string Name { get; init; } = "";
+
+    [JsonPropertyName("size")]
+    public long? Size { get; init; }
+
+    [JsonPropertyName("required")]
+    public bool Required { get; init; }
+}
diff --git a/native-app/src/VoxType.NativeCore/ModelManager.cs b/native-app/src/VoxType.NativeCore/ModelManager.cs
new file mode 100644
index 0000000..68f2c68
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/ModelManager.cs
@@ -0,0 +1,206 @@
+using System.Net;
+using System.Net.Http.Headers;
+
+namespace VoxType.NativeCore;
+
+public sealed record ModelDownloadProgress(double Percentage, string Message);
+
+public sealed class ModelManager : IDisposable
+{
+    private readonly HttpClient _httpClient;
+
+    public ModelManager()
+    {
+        _httpClient = new HttpClient(new HttpClientHandler
+        {
+            AutomaticDecompression = DecompressionMethods.All,
+        })
+        {
+            Timeout = TimeSpan.FromMinutes(30),
+        };
+        _httpClient.DefaultRequestHeaders.UserAgent.ParseAdd("VoxType-NativePreview/0.1");
+    }
+
+    public static bool IsInstalled(ModelCatalogEntry model)
+    {
+        var directory = VoxTypePaths.ResolveModelDirectory(model);
+        var requiredFiles = model.Download.ModelScopeFiles.Where(file => file.Required).ToArray();
+        return requiredFiles.Length > 0
+               && requiredFiles.All(file =>
+               {
+                   ValidateFileName(file.Name);
+                   return IsComplete(Path.Combine(directory, file.Name), file.Size);
+               });
+    }
+
+    public async Task<string> EnsureInstalledAsync(
+        ModelCatalogEntry model,
+        IProgress<ModelDownloadProgress>? progress = null,
+        CancellationToken cancellationToken = default)
+    {
+        if (string.IsNullOrWhiteSpace(model.Download.ModelScopeResolveBase))
+        {
+            throw new InvalidDataException($"模型 {model.Id} 缺少 ModelScope 下载地址。");
+        }
+
+        var requiredFiles = model.Download.ModelScopeFiles.Where(file => file.Required).ToArray();
+        if (requiredFiles.Length == 0)
+        {
+            throw new InvalidDataException($"模型 {model.Id} 未配置必需文件。");
+        }
+
+        var directory = VoxTypePaths.ResolveModelDirectory(model);
+        Directory.CreateDirectory(directory);
+
+        var knownTotal = requiredFiles.Sum(file => file.Size ?? 0L);
+        long completedBytes = 0;
+
+        foreach (var file in requiredFiles)
+        {
+            cancellationToken.ThrowIfCancellationRequested();
+            ValidateFileName(file.Name);
+            var destination = Path.Combine(directory, file.Name);
+            if (IsComplete(destination, file.Size))
+            {
+                completedBytes += file.Size ?? new FileInfo(destination).Length;
+                Report(progress, completedBytes, knownTotal, $"已存在：{file.Name}");
+                continue;
+            }
+
+            var url = $"{model.Download.ModelScopeResolveBase.TrimEnd('/')}/{Uri.EscapeDataString(file.Name)}";
+            await DownloadFileAsync(
+                    url,
+                    destination,
+                    file.Size,
+                    completedBytes,
+                    knownTotal,
+                    progress,
+                    cancellationToken)
+                .ConfigureAwait(false);
+
+            completedBytes += file.Size ?? new FileInfo(destination).Length;
+        }
+
+        if (!IsInstalled(model))
+        {
+            throw new InvalidDataException("模型下载结束，但完整性检查未通过。");
+        }
+
+        progress?.Report(new ModelDownloadProgress(100, "模型已就绪"));
+        return directory;
+    }
+
+    private async Task DownloadFileAsync(
+        string url,
+        string destination,
+        long? expectedSize,
+        long completedBeforeFile,
+        long knownTotal,
+        IProgress<ModelDownloadProgress>? progress,
+        CancellationToken cancellationToken)
+    {
+        var partPath = destination + ".part";
+        var existing = File.Exists(partPath) ? new FileInfo(partPath).Length : 0L;
+        if (expectedSize is > 0 && existing > expectedSize)
+        {
+            File.Delete(partPath);
+            existing = 0;
+        }
+
+        using var request = new HttpRequestMessage(HttpMethod.Get, url);
+        if (existing > 0)
+        {
+            request.Headers.Range = new RangeHeaderValue(existing, null);
+        }
+
+        using var response = await _httpClient.SendAsync(
+                request,
+                HttpCompletionOption.ResponseHeadersRead,
+                cancellationToken)
+            .ConfigureAwait(false);
+
+        if (existing > 0 && response.StatusCode != HttpStatusCode.PartialContent)
+        {
+            existing = 0;
+        }
+
+        response.EnsureSuccessStatusCode();
+
+        await using var input = await response.Content.ReadAsStreamAsync(cancellationToken).ConfigureAwait(false);
+        await using var output = new FileStream(
+            partPath,
+            existing > 0 ? FileMode.Append : FileMode.Create,
+            FileAccess.Write,
+            FileShare.Read,
+            1024 * 128,
+            FileOptions.Asynchronous | FileOptions.SequentialScan);
+
+        var buffer = new byte[1024 * 128];
+        var downloaded = existing;
+        var lastReported = DateTime.UtcNow;
+        while (true)
+        {
+            var count = await input.ReadAsync(buffer, cancellationToken).ConfigureAwait(false);
+            if (count == 0)
+            {
+                break;
+            }
+
+            await output.WriteAsync(buffer.AsMemory(0, count), cancellationToken).ConfigureAwait(false);
+            downloaded += count;
+
+            if (DateTime.UtcNow - lastReported >= TimeSpan.FromMilliseconds(150))
+            {
+                Report(
+                    progress,
+                    completedBeforeFile + downloaded,
+                    knownTotal,
+                    $"下载 {Path.GetFileName(destination)} · {downloaded / 1024d / 1024d:F1} MB");
+                lastReported = DateTime.UtcNow;
+            }
+        }
+
+        await output.FlushAsync(cancellationToken).ConfigureAwait(false);
+        if (expectedSize is > 0 && downloaded != expectedSize)
+        {
+            throw new InvalidDataException(
+                $"{Path.GetFileName(destination)} 下载不完整：{downloaded}/{expectedSize} bytes");
+        }
+
+        File.Move(partPath, destination, true);
+    }
+
+    private static bool IsComplete(string path, long? expectedSize)
+    {
+        if (!File.Exists(path))
+        {
+            return false;
+        }
+
+        var length = new FileInfo(path).Length;
+        return expectedSize is > 0 ? length == expectedSize : length > 0;
+    }
+
+    private static void ValidateFileName(string fileName)
+    {
+        if (string.IsNullOrWhiteSpace(fileName)
+            || fileName is "." or ".."
+            || !string.Equals(Path.GetFileName(fileName), fileName, StringComparison.Ordinal)
+            || fileName.IndexOfAny(Path.GetInvalidFileNameChars()) >= 0)
+        {
+            throw new InvalidDataException($"模型文件名无效：{fileName}");
+        }
+    }
+
+    private static void Report(
+        IProgress<ModelDownloadProgress>? progress,
+        long current,
+        long total,
+        string message)
+    {
+        var percentage = total > 0 ? Math.Clamp(current * 100d / total, 0, 100) : 0;
+        progress?.Report(new ModelDownloadProgress(percentage, message));
+    }
+
+    public void Dispose() => _httpClient.Dispose();
+}
diff --git a/native-app/src/VoxType.NativeCore/SenseVoiceEngine.cs b/native-app/src/VoxType.NativeCore/SenseVoiceEngine.cs
new file mode 100644
index 0000000..ed7de7b
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/SenseVoiceEngine.cs
@@ -0,0 +1,124 @@
+using SherpaOnnx;
+
+namespace VoxType.NativeCore;
+
+public sealed class SenseVoiceEngine : IDisposable
+{
+    private readonly SemaphoreSlim _operationGate = new(1, 1);
+    private OfflineRecognizer? _recognizer;
+    private string? _loadedModelDirectory;
+    private bool _disposed;
+
+    public async Task WarmUpAsync(string modelDirectory, CancellationToken cancellationToken = default)
+    {
+        await _operationGate.WaitAsync(cancellationToken).ConfigureAwait(false);
+        try
+        {
+            ThrowIfDisposed();
+            await EnsureLoadedCoreAsync(modelDirectory, cancellationToken).ConfigureAwait(false);
+        }
+        finally
+        {
+            _operationGate.Release();
+        }
+    }
+
+    public async Task<string> TranscribeAsync(
+        string modelDirectory,
+        float[] samples,
+        CancellationToken cancellationToken = default)
+    {
+        if (samples.Length == 0)
+        {
+            return "";
+        }
+
+        await _operationGate.WaitAsync(cancellationToken).ConfigureAwait(false);
+        try
+        {
+            ThrowIfDisposed();
+            await EnsureLoadedCoreAsync(modelDirectory, cancellationToken).ConfigureAwait(false);
+            cancellationToken.ThrowIfCancellationRequested();
+
+            var recognizer = _recognizer ?? throw new InvalidOperationException("识别器尚未加载。");
+            return await Task.Run(() =>
+            {
+                using var stream = recognizer.CreateStream();
+                stream.AcceptWaveform(AudioCaptureSession.SampleRate, samples);
+                recognizer.Decode(stream);
+                return stream.Result.Text.Trim();
+            }, cancellationToken).ConfigureAwait(false);
+        }
+        finally
+        {
+            _operationGate.Release();
+        }
+    }
+
+    private async Task EnsureLoadedCoreAsync(string modelDirectory, CancellationToken cancellationToken)
+    {
+        var fullDirectory = Path.GetFullPath(modelDirectory);
+        if (_recognizer is not null
+            && string.Equals(_loadedModelDirectory, fullDirectory, StringComparison.OrdinalIgnoreCase))
+        {
+            return;
+        }
+
+        var modelPath = Path.Combine(fullDirectory, "model.int8.onnx");
+        if (!File.Exists(modelPath))
+        {
+            modelPath = Path.Combine(fullDirectory, "model.onnx");
+        }
+
+        var tokensPath = Path.Combine(fullDirectory, "tokens.txt");
+        if (!File.Exists(modelPath) || !File.Exists(tokensPath))
+        {
+            throw new FileNotFoundException("SenseVoice 模型文件不完整。", fullDirectory);
+        }
+
+        var config = new OfflineRecognizerConfig();
+        config.FeatConfig.SampleRate = AudioCaptureSession.SampleRate;
+        config.FeatConfig.FeatureDim = 80;
+        config.ModelConfig.Tokens = tokensPath;
+        config.ModelConfig.NumThreads = Math.Clamp(Environment.ProcessorCount / 2, 1, 4);
+        config.ModelConfig.Provider = "cpu";
+        config.ModelConfig.Debug = 0;
+        config.ModelConfig.SenseVoice.Model = modelPath;
+        config.ModelConfig.SenseVoice.Language = "auto";
+        config.ModelConfig.SenseVoice.UseInverseTextNormalization = 1;
+        config.DecodingMethod = "greedy_search";
+
+        var next = await Task.Run(() => new OfflineRecognizer(config), cancellationToken).ConfigureAwait(false);
+        var previous = _recognizer;
+        _recognizer = next;
+        _loadedModelDirectory = fullDirectory;
+        previous?.Dispose();
+    }
+
+    private void ThrowIfDisposed()
+    {
+        ObjectDisposedException.ThrowIf(_disposed, this);
+    }
+
+    public void Dispose()
+    {
+        if (_disposed)
+        {
+            return;
+        }
+
+        _disposed = true;
+        // Native decode cannot be cancelled midway. Wait for an active decode before releasing
+        // the recognizer so closing the preview cannot race a P/Invoke call.
+        _operationGate.Wait();
+        try
+        {
+            _recognizer?.Dispose();
+            _recognizer = null;
+        }
+        finally
+        {
+            _operationGate.Release();
+        }
+    }
+}
diff --git a/native-app/src/VoxType.NativeCore/VoxType.NativeCore.csproj b/native-app/src/VoxType.NativeCore/VoxType.NativeCore.csproj
new file mode 100644
index 0000000..e2dd24a
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/VoxType.NativeCore.csproj
@@ -0,0 +1,14 @@
+<Project Sdk="Microsoft.NET.Sdk">
+  <PropertyGroup>
+    <TargetFramework>net10.0-windows</TargetFramework>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <Nullable>enable</Nullable>
+    <PlatformTarget>x64</PlatformTarget>
+    <RuntimeIdentifier>win-x64</RuntimeIdentifier>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <PackageReference Include="NAudio.WinMM" />
+    <PackageReference Include="org.k2fsa.sherpa.onnx" />
+  </ItemGroup>
+</Project>
diff --git a/native-app/src/VoxType.NativeCore/VoxTypePaths.cs b/native-app/src/VoxType.NativeCore/VoxTypePaths.cs
new file mode 100644
index 0000000..0b90543
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/VoxTypePaths.cs
@@ -0,0 +1,37 @@
+namespace VoxType.NativeCore;
+
+public static class VoxTypePaths
+{
+    public static string DataRoot
+    {
+        get
+        {
+            var custom = Environment.GetEnvironmentVariable("VOXTYPE_DATA_ROOT");
+            if (!string.IsNullOrWhiteSpace(custom))
+            {
+                return Path.GetFullPath(custom);
+            }
+
+            return Path.Combine(
+                Environment.GetFolderPath(Environment.SpecialFolder.LocalApplicationData),
+                "VoxType");
+        }
+    }
+
+    public static string ModelsRoot => Path.Combine(DataRoot, "models");
+
+    public static string ResolveModelDirectory(ModelCatalogEntry model)
+    {
+        if (string.IsNullOrWhiteSpace(model.Layout)
+            || model.Layout is "." or ".."
+            || !string.Equals(Path.GetFileName(model.Layout), model.Layout, StringComparison.Ordinal)
+            || model.Layout.IndexOfAny(Path.GetInvalidFileNameChars()) >= 0
+            || model.Layout.Contains(Path.DirectorySeparatorChar)
+            || model.Layout.Contains(Path.AltDirectorySeparatorChar))
+        {
+            throw new InvalidDataException($"模型 layout 无效：{model.Layout}");
+        }
+
+        return Path.Combine(ModelsRoot, model.Layout);
+    }
+}
diff --git a/native-app/src/VoxType.NativeCore/WaveFileAudio.cs b/native-app/src/VoxType.NativeCore/WaveFileAudio.cs
new file mode 100644
index 0000000..f3d677d
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/WaveFileAudio.cs
@@ -0,0 +1,42 @@
+using NAudio.Wave;
+using NAudio.Wave.SampleProviders;
+
+namespace VoxType.NativeCore;
+
+public static class WaveFileAudio
+{
+    public static float[] ReadMono16Khz(string path)
+    {
+        using var reader = new WaveFileReader(path);
+        ISampleProvider provider = reader.ToSampleProvider();
+
+        if (provider.WaveFormat.Channels == 2)
+        {
+            provider = new StereoToMonoSampleProvider(provider);
+        }
+        else if (provider.WaveFormat.Channels != 1)
+        {
+            throw new NotSupportedException($"仅支持单声道或双声道音频，当前为 {provider.WaveFormat.Channels} 声道。");
+        }
+
+        if (provider.WaveFormat.SampleRate != AudioCaptureSession.SampleRate)
+        {
+            provider = new WdlResamplingSampleProvider(provider, AudioCaptureSession.SampleRate);
+        }
+
+        var samples = new List<float>();
+        var buffer = new float[AudioCaptureSession.SampleRate];
+        while (true)
+        {
+            var count = provider.Read(buffer, 0, buffer.Length);
+            if (count == 0)
+            {
+                break;
+            }
+
+            samples.AddRange(buffer.AsSpan(0, count).ToArray());
+        }
+
+        return [.. samples];
+    }
+}
diff --git a/native-app/src/VoxType.NativeSmoke/Program.cs b/native-app/src/VoxType.NativeSmoke/Program.cs
new file mode 100644
index 0000000..1d1f13c
--- /dev/null
+++ b/native-app/src/VoxType.NativeSmoke/Program.cs
@@ -0,0 +1,56 @@
+using System.Diagnostics;
+using VoxType.NativeCore;
+
+if (args.Length == 0 || args.Contains("--help", StringComparer.OrdinalIgnoreCase))
+{
+    Console.WriteLine("Usage: VoxType.NativeSmoke <audio.wav> [model-directory] [repeat-count]");
+    Console.WriteLine($"Default model directory: {Path.Combine(VoxTypePaths.ModelsRoot, "sensevoice")}");
+    return 0;
+}
+
+var audioPath = Path.GetFullPath(args[0]);
+var modelDirectory = args.Length > 1
+    ? Path.GetFullPath(args[1])
+    : Path.Combine(VoxTypePaths.ModelsRoot, "sensevoice");
+var repeatCount = args.Length > 2 && int.TryParse(args[2], out var parsedRepeatCount)
+    ? Math.Clamp(parsedRepeatCount, 1, 20)
+    : 1;
+
+if (!File.Exists(audioPath))
+{
+    Console.Error.WriteLine($"Audio file not found: {audioPath}");
+    return 2;
+}
+
+try
+{
+    var samples = WaveFileAudio.ReadMono16Khz(audioPath);
+    var audioSeconds = samples.Length / (double)AudioCaptureSession.SampleRate;
+    Console.WriteLine($"Audio: {audioSeconds:F2}s, samples={samples.Length}");
+    Console.WriteLine($"Model: {modelDirectory}");
+
+    using var engine = new SenseVoiceEngine();
+    var loadWatch = Stopwatch.StartNew();
+    await engine.WarmUpAsync(modelDirectory);
+    loadWatch.Stop();
+    Console.WriteLine($"Model load: {loadWatch.Elapsed.TotalSeconds:F2}s");
+
+    var hasText = false;
+    for (var index = 1; index <= repeatCount; index++)
+    {
+        var watch = Stopwatch.StartNew();
+        var text = await engine.TranscribeAsync(modelDirectory, samples);
+        watch.Stop();
+
+        Console.WriteLine($"Run {index}: {watch.Elapsed.TotalSeconds:F2}s, RTF={watch.Elapsed.TotalSeconds / audioSeconds:F3}");
+        Console.WriteLine($"Text: {text}");
+        hasText |= !string.IsNullOrWhiteSpace(text);
+    }
+
+    return hasText ? 0 : 3;
+}
+catch (Exception ex)
+{
+    Console.Error.WriteLine(ex);
+    return 1;
+}
diff --git a/native-app/src/VoxType.NativeSmoke/VoxType.NativeSmoke.csproj b/native-app/src/VoxType.NativeSmoke/VoxType.NativeSmoke.csproj
new file mode 100644
index 0000000..7e6177b
--- /dev/null
+++ b/native-app/src/VoxType.NativeSmoke/VoxType.NativeSmoke.csproj
@@ -0,0 +1,14 @@
+<Project Sdk="Microsoft.NET.Sdk">
+  <PropertyGroup>
+    <OutputType>Exe</OutputType>
+    <TargetFramework>net10.0-windows</TargetFramework>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <Nullable>enable</Nullable>
+    <PlatformTarget>x64</PlatformTarget>
+    <RuntimeIdentifier>win-x64</RuntimeIdentifier>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <ProjectReference Include="../VoxType.NativeCore/VoxType.NativeCore.csproj" />
+  </ItemGroup>
+</Project>

From 0cb908e0daa6f04cc6c94a1d2b641cbbf7619f2b Mon Sep 17 00:00:00 2001
From: Cea <61349137+ceastld@users.noreply.github.com>
Date: Fri, 7 Aug 2026 19:49:38 +0800
Subject: [PATCH 2/4] ci: update native preview actions

---
 .github/workflows/native-preview.yml | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/native-preview.yml b/.github/workflows/native-preview.yml
index 3ec1d80..1ee64f5 100644
--- a/.github/workflows/native-preview.yml
+++ b/.github/workflows/native-preview.yml
@@ -22,10 +22,10 @@ jobs:
     runs-on: windows-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v7
 
       - name: Setup .NET 10
-        uses: actions/setup-dotnet@v4
+        uses: actions/setup-dotnet@v6
         with:
           dotnet-version: '10.0.x'
 
@@ -38,7 +38,7 @@ jobs:
         run: pwsh -NoProfile -File ./native-app/build.ps1 -SkipBuild
 
       - name: Upload preview zip
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v7
         with:
           name: VoxType-NativePreview-win-x64
           path: native-app/publish/VoxType.NativePreview-win-x64.zip

From b191312569438da0ff3c8f2793ed2d0adafecc4a Mon Sep 17 00:00:00 2001
From: Cea <61349137+ceastld@users.noreply.github.com>
Date: Fri, 7 Aug 2026 21:02:36 +0800
Subject: [PATCH 3/4] feat: stream native speech recognition

---
 docs/quicker-integration-research.md          |   2 +-
 native-app/BENCHMARKS.md                      |  28 +-
 native-app/README.md                          |  27 +-
 native-app/catalog/models.json                |  34 ++
 .../src/VoxType.NativeApp/MainWindow.xaml     |  11 +-
 .../src/VoxType.NativeApp/MainWindow.xaml.cs  |  94 +++--
 .../VoxType.NativeApp.csproj                  |   2 +-
 .../VoxType.NativeCore/AudioCaptureSession.cs |  39 +-
 .../src/VoxType.NativeCore/ModelCatalog.cs    |  14 +
 .../src/VoxType.NativeCore/ModelManager.cs    |  32 +-
 .../StreamingVoiceEngine.cs                   | 337 ++++++++++++++++++
 native-app/src/VoxType.NativeSmoke/Program.cs |  44 ++-
 12 files changed, 580 insertions(+), 84 deletions(-)
 create mode 100644 native-app/catalog/models.json
 create mode 100644 native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs

diff --git a/docs/quicker-integration-research.md b/docs/quicker-integration-research.md
index 77a6369..a3ea75c 100644
--- a/docs/quicker-integration-research.md
+++ b/docs/quicker-integration-research.md
@@ -17,7 +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，用于验证更小的原生 worker；它不改变“Quicker 主进程不直接承载模型”的隔离边界。
+- `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
 
diff --git a/native-app/BENCHMARKS.md b/native-app/BENCHMARKS.md
index 50ffcaf..f858f96 100644
--- a/native-app/BENCHMARKS.md
+++ b/native-app/BENCHMARKS.md
@@ -7,29 +7,30 @@
 - 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，CPU provider，最多 4 线程
-- 模型：SenseVoice Small INT8，239,233,841-byte ONNX
+- 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 | 1.27–1.63 s |
-| 稳态识别 | 0.16–0.23 s，RTF 0.032–0.045 |
+| 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 | 8 次文字更新，首个 partial 约 0.91 s |
 | 发布目录 | 21.4 MB |
 | zip | 8.1 MB |
-| 模型加载后的进程 Working Set | 约 506 MB |
-| 模型加载后的 Private Memory | 约 450 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。重复运行存在正常调度波动，因此表中保留两轮实测范围。
+`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
 
-- 首次加载和后续识别必须分开测量；UI 启动时预热模型，因此用户松开录音后的主要等待接近“稳态识别”。
-- 体积下降主要来自移除 Python、NumPy、aiohttp 和 PyInstaller runtime；发布目录仍主要由 `onnxruntime.dll` 与 `sherpa-onnx-c-api.dll` 构成。
-- 内存主要由 229 MiB 模型、ONNX Runtime 权重/工作区和 WPF 进程构成。后续 Quicker 接入仍建议使用独立原生 worker，以隔离这部分内存和 native DLL 生命周期。
-- 当前只启用 CPU provider。DirectML/CUDA 是否值得加入，要以更多硬件上的启动失败率、驱动兼容和安装包增量为准。
+- `--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。
+- 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
 
@@ -38,6 +39,7 @@ pwsh ./build.ps1
 
 dotnet run --project ./src/VoxType.NativeSmoke -c Release -- `
   <audio.wav> `
-  "$env:LOCALAPPDATA\VoxType\models\sensevoice" `
-  3
+  "$env:LOCALAPPDATA\VoxType\models\zipformer-zh-streaming-small" `
+  1 `
+  --realtime
 ```
diff --git a/native-app/README.md b/native-app/README.md
index 87bafe8..78cb00c 100644
--- a/native-app/README.md
+++ b/native-app/README.md
@@ -1,6 +1,6 @@
 # VoxType Native Preview
 
-面向 `QuickerPc/Quicker` 的最小原生语音识别体验版。它使用 `.NET 10 + WPF + NAudio + sherpa-onnx C#`，直接加载 SenseVoice，不启动 Python、aiohttp 或 PyInstaller runtime。
+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
 
@@ -9,13 +9,13 @@ cd native-app
 dotnet run --project ./src/VoxType.NativeApp -c Release
 ```
 
-首次启动点击“下载模型”。模型约 229 MB，保存到 `%LOCALAPPDATA%\VoxType\models\sensevoice`，与现有 VoxType 共用，不进入应用安装包。模型就绪后：
+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“开始录音”。
-2. 说话后点击“结束并识别”。
-3. 查看耗时和 RTF，复制识别结果。
+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。
 
-当前预览版只做录音、识别和复制，不注册全局热键、不自动向其它窗口输入文字。这让首轮测试集中在原生运行稳定性、识别性能和打包体积。
+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
 
@@ -28,7 +28,7 @@ pwsh ./build.ps1
 - `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。
 
-首轮实测 zip 为 8.1 MB，模型加载后的 5 秒中文音频稳态识别为 0.16–0.23 秒。环境、内存与复现命令见 [`BENCHMARKS.md`](BENCHMARKS.md)。
+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`。
 
@@ -37,14 +37,15 @@ pwsh ./build.ps1
 ## 命令行识别验证
 
 ```powershell
-dotnet run --project ./src/VoxType.NativeSmoke -c Release -- <audio.wav> [model-directory] [repeat-count]
+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，用于区分模型加载耗时和稳态 decode 耗时。该工具用于 CI/开发验证，不进入预览应用压缩包。
+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
 
-- 只启用 Windows x64 CPU provider，优先保证部署一致性；GPU provider 留待实测收益与驱动兼容性后再开放。
-- 识别器在进程内只加载一次，每次录音创建独立 stream，避免重复加载模型。
-- 模型按 catalog 下载并支持 `.part` 断点续传。
-- 原生识别调用不可被 sherpa-onnx 中途打断；关闭窗口会阻止新会话，但正在执行的单次 decode 需要自然返回。
+- 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。
+- 端点检测在连续静音后确认当前分段并继续同一录音会话；点击停止时额外送入右侧静音上下文以刷新最终结果。
+- 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。
diff --git a/native-app/catalog/models.json b/native-app/catalog/models.json
new file mode 100644
index 0000000..d507b2f
--- /dev/null
+++ b/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"
+          }
+        ]
+      }
+    }
+  ]
+}
diff --git a/native-app/src/VoxType.NativeApp/MainWindow.xaml b/native-app/src/VoxType.NativeApp/MainWindow.xaml
index d4d7142..984a325 100644
--- a/native-app/src/VoxType.NativeApp/MainWindow.xaml
+++ b/native-app/src/VoxType.NativeApp/MainWindow.xaml
@@ -54,7 +54,10 @@
                     <ColumnDefinition Width="Auto" />
                 </Grid.ColumnDefinitions>
                 <StackPanel>
-                    <TextBlock FontSize="15" FontWeight="SemiBold" Text="SenseVoice Small INT8" />
+                    <TextBlock x:Name="ModelNameText"
+                               FontSize="15"
+                               FontWeight="SemiBold"
+                               Text="中文流式模型" />
                     <TextBlock x:Name="ModelStatusText"
                                Margin="0,5,0,0"
                                Foreground="{StaticResource MutedTextBrush}"
@@ -96,13 +99,13 @@
                         Height="54"
                         IsEnabled="False"
                         Click="RecordButton_Click"
-                        Content="开始录音" />
+                        Content="开始实时识别" />
                 <StackPanel Grid.Column="1" Margin="18,0" VerticalAlignment="Center">
                     <TextBlock x:Name="SessionTitleText" FontSize="15" FontWeight="SemiBold" Text="准备就绪" />
                     <TextBlock x:Name="SessionHintText"
                                Margin="0,4,0,0"
                                Foreground="{StaticResource MutedTextBrush}"
-                               Text="点击开始录音，再次点击结束并识别" />
+                               Text="点击开始，结果会随着说话实时刷新" />
                 </StackPanel>
                 <TextBlock x:Name="DurationText"
                            Grid.Column="2"
@@ -128,7 +131,7 @@
                     <RowDefinition Height="Auto" />
                 </Grid.RowDefinitions>
                 <Grid>
-                    <TextBlock FontSize="14" FontWeight="SemiBold" Text="识别结果" />
+                    <TextBlock FontSize="14" FontWeight="SemiBold" Text="实时识别结果" />
                     <StackPanel HorizontalAlignment="Right" Orientation="Horizontal">
                         <Button x:Name="ClearButton"
                                 Margin="0,0,8,0"
diff --git a/native-app/src/VoxType.NativeApp/MainWindow.xaml.cs b/native-app/src/VoxType.NativeApp/MainWindow.xaml.cs
index 3df69d5..a88cb9e 100644
--- a/native-app/src/VoxType.NativeApp/MainWindow.xaml.cs
+++ b/native-app/src/VoxType.NativeApp/MainWindow.xaml.cs
@@ -10,7 +10,7 @@ namespace VoxType.NativeApp;
 
 public partial class MainWindow : Window
 {
-    private readonly SenseVoiceEngine _engine = new();
+    private readonly StreamingVoiceEngine _engine = new();
     private readonly ModelManager _modelManager = new();
     private readonly CancellationTokenSource _lifetime = new();
     private readonly DispatcherTimer _recordingTimer;
@@ -34,7 +34,8 @@ private async void Window_Loaded(object sender, RoutedEventArgs e)
         try
         {
             var catalogPath = Path.Combine(AppContext.BaseDirectory, "catalog", "models.json");
-            _model = ModelCatalog.Load(catalogPath).GetDefaultSenseVoice();
+            _model = ModelCatalog.Load(catalogPath).GetDefaultStreamingModel();
+            ModelNameText.Text = _model.Name;
             RefreshModelState();
             if (ModelManager.IsInstalled(_model))
             {
@@ -42,7 +43,7 @@ private async void Window_Loaded(object sender, RoutedEventArgs e)
             }
             else
             {
-                StatusText.Text = "首次使用需下载约 229 MB 的模型文件，模型不会打进应用包。";
+                StatusText.Text = "首次使用需下载约 25 MB 的中文流式模型，模型不会打进应用包。";
             }
         }
         catch (Exception ex)
@@ -98,13 +99,13 @@ private async Task WarmUpAsync()
         SetBusy(true);
         SessionTitleText.Text = "正在加载模型";
         SessionHintText.Text = "首次加载通常需要几秒";
-        StatusText.Text = "正在创建原生 SenseVoice 识别器…";
+        StatusText.Text = "正在创建原生 Zipformer 在线识别器…";
         var watch = Stopwatch.StartNew();
         try
         {
             await _engine.WarmUpAsync(VoxTypePaths.ResolveModelDirectory(_model), _lifetime.Token);
             SessionTitleText.Text = "准备就绪";
-            SessionHintText.Text = "点击开始录音，再次点击结束并识别";
+            SessionHintText.Text = "点击开始，结果会随着说话实时刷新";
             StatusText.Text = $"模型已加载 · {watch.Elapsed.TotalSeconds:F2}s";
         }
         finally
@@ -122,34 +123,45 @@ private async void RecordButton_Click(object sender, RoutedEventArgs e)
 
         if (_capture is null)
         {
-            StartRecording();
+            await StartRecordingAsync();
         }
         else
         {
-            await StopAndTranscribeAsync();
+            await StopStreamingAsync();
         }
     }
 
-    private void StartRecording()
+    private async Task StartRecordingAsync()
     {
         try
         {
-            _capture = AudioCaptureSession.Start();
+            _engine.StartSession(OnStreamingUpdate);
+            _capture = AudioCaptureSession.Start((_, args) => _engine.AcceptAudio(args.Samples));
             _recordingWatch = Stopwatch.StartNew();
             _recordingTimer.Start();
-            RecordButton.Content = "结束并识别";
+            RecordButton.Content = "停止识别";
             RecordButton.Background = new SolidColorBrush(Color.FromRgb(213, 75, 75));
-            SessionTitleText.Text = "正在录音";
-            SessionHintText.Text = "自然说话，结束后会在本机识别";
-            StatusText.Text = "麦克风录音中…";
+            SessionTitleText.Text = "正在实时识别";
+            SessionHintText.Text = "自然说话，临时结果会持续更新";
+            StatusText.Text = "麦克风录音与本地流式识别进行中…";
+            MetricsText.Text = "等待第一段语音…";
         }
         catch (Exception ex)
         {
+            try
+            {
+                await _engine.StopSessionAsync();
+            }
+            catch
+            {
+                // The session may not have started; preserve the microphone error below.
+            }
+
             ShowError("无法开始录音", ex);
         }
     }
 
-    private async Task StopAndTranscribeAsync()
+    private async Task StopStreamingAsync()
     {
         var capture = _capture;
         _capture = null;
@@ -161,35 +173,29 @@ private async Task StopAndTranscribeAsync()
         _recordingTimer.Stop();
         _recordingWatch?.Stop();
         SetBusy(true);
-        RecordButton.Content = "开始录音";
+        RecordButton.Content = "开始实时识别";
         RecordButton.ClearValue(BackgroundProperty);
-        SessionTitleText.Text = "正在识别";
-        SessionHintText.Text = "原生推理不会上传音频";
-        StatusText.Text = "正在运行 SenseVoice…";
+        SessionTitleText.Text = "正在提交最终结果";
+        SessionHintText.Text = "正在刷新最后一段上下文";
+        StatusText.Text = "正在结束本地流式会话…";
 
         try
         {
-            var samples = await capture.StopAsync(_lifetime.Token);
+            await capture.StopAsync(_lifetime.Token);
             capture.Dispose();
-            var audioSeconds = samples.Length / (double)AudioCaptureSession.SampleRate;
-            if (audioSeconds < 0.25)
+            var result = await _engine.StopSessionAsync();
+            if (result.AudioSeconds < 0.25)
             {
                 throw new InvalidOperationException("录音太短，请至少说 0.3 秒。");
             }
 
-            var recognitionWatch = Stopwatch.StartNew();
-            var text = await _engine.TranscribeAsync(
-                VoxTypePaths.ResolveModelDirectory(_model),
-                samples,
-                _lifetime.Token);
-            recognitionWatch.Stop();
-
-            ResultTextBox.Text = text;
-            CopyButton.IsEnabled = !string.IsNullOrWhiteSpace(text);
-            MetricsText.Text = $"音频 {audioSeconds:F1}s · 识别 {recognitionWatch.Elapsed.TotalSeconds:F2}s · RTF {recognitionWatch.Elapsed.TotalSeconds / audioSeconds:F2}";
-            SessionTitleText.Text = string.IsNullOrWhiteSpace(text) ? "未识别到文字" : "识别完成";
-            SessionHintText.Text = "可复制结果，或继续下一段录音";
-            StatusText.Text = string.IsNullOrWhiteSpace(text) ? "没有识别结果，请靠近麦克风后重试。" : "识别结果已生成。";
+            ResultTextBox.Text = result.Text;
+            CopyButton.IsEnabled = !string.IsNullOrWhiteSpace(result.Text);
+            var firstText = result.FirstTextSeconds is null ? "无" : $"{result.FirstTextSeconds:F2}s";
+            MetricsText.Text = $"音频 {result.AudioSeconds:F1}s · 首字 {firstText} · 会话 {result.ElapsedSeconds:F2}s";
+            SessionTitleText.Text = string.IsNullOrWhiteSpace(result.Text) ? "未识别到文字" : "实时识别完成";
+            SessionHintText.Text = "可复制结果，或继续下一段实时识别";
+            StatusText.Text = string.IsNullOrWhiteSpace(result.Text) ? "没有识别结果，请靠近麦克风后重试。" : "最终结果已提交。";
         }
         catch (OperationCanceledException)
         {
@@ -206,6 +212,24 @@ private async Task StopAndTranscribeAsync()
         }
     }
 
+    private void OnStreamingUpdate(StreamingRecognitionUpdate update)
+    {
+        _ = Dispatcher.InvokeAsync(() =>
+        {
+            if (_lifetime.IsCancellationRequested)
+            {
+                return;
+            }
+
+            ResultTextBox.Text = update.Text;
+            ResultTextBox.CaretIndex = ResultTextBox.Text.Length;
+            ResultTextBox.ScrollToEnd();
+            CopyButton.IsEnabled = !string.IsNullOrWhiteSpace(update.Text);
+            MetricsText.Text = $"已接收 {update.AudioSeconds:F1}s 音频 · {update.ElapsedSeconds:F1}s 实时会话";
+            StatusText.Text = update.IsEndpoint ? "已确认一段，继续聆听…" : "正在实时识别…";
+        }, DispatcherPriority.Background);
+    }
+
     private void CopyButton_Click(object sender, RoutedEventArgs e)
     {
         if (string.IsNullOrWhiteSpace(ResultTextBox.Text))
@@ -245,7 +269,7 @@ private void RefreshModelState()
         var installed = ModelManager.IsInstalled(_model);
         ModelStatusText.Text = installed
             ? $"已安装 · {VoxTypePaths.ResolveModelDirectory(_model)}"
-            : "未安装 · 首次使用下载约 229 MB";
+            : "未安装 · 首次使用下载约 25 MB";
         DownloadModelButton.Content = installed ? "校验/续传" : "下载模型";
         RecordButton.IsEnabled = installed && !_busy;
     }
diff --git a/native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj b/native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj
index 19af92d..525e027 100644
--- a/native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj
+++ b/native-app/src/VoxType.NativeApp/VoxType.NativeApp.csproj
@@ -14,7 +14,7 @@
 
   <ItemGroup>
     <ProjectReference Include="../VoxType.NativeCore/VoxType.NativeCore.csproj" />
-    <Content Include="../../../catalog/models.json" Link="catalog/models.json">
+    <Content Include="../../catalog/models.json" Link="catalog/models.json">
       <CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
       <CopyToPublishDirectory>PreserveNewest</CopyToPublishDirectory>
     </Content>
diff --git a/native-app/src/VoxType.NativeCore/AudioCaptureSession.cs b/native-app/src/VoxType.NativeCore/AudioCaptureSession.cs
index 547d78c..90034d5 100644
--- a/native-app/src/VoxType.NativeCore/AudioCaptureSession.cs
+++ b/native-app/src/VoxType.NativeCore/AudioCaptureSession.cs
@@ -2,6 +2,11 @@
 
 namespace VoxType.NativeCore;
 
+public sealed class AudioSamplesAvailableEventArgs(float[] samples) : EventArgs
+{
+    public float[] Samples { get; } = samples;
+}
+
 public sealed class AudioCaptureSession : IDisposable
 {
     public const int SampleRate = 16_000;
@@ -12,6 +17,8 @@ public sealed class AudioCaptureSession : IDisposable
     private readonly TaskCompletionSource _stopped = new(TaskCreationOptions.RunContinuationsAsynchronously);
     private bool _isStopped;
 
+    public event EventHandler<AudioSamplesAvailableEventArgs>? SamplesAvailable;
+
     private AudioCaptureSession(WaveInEvent waveIn)
     {
         _waveIn = waveIn;
@@ -19,7 +26,8 @@ private AudioCaptureSession(WaveInEvent waveIn)
         _waveIn.RecordingStopped += OnRecordingStopped;
     }
 
-    public static AudioCaptureSession Start()
+    public static AudioCaptureSession Start(
+        EventHandler<AudioSamplesAvailableEventArgs>? samplesAvailable = null)
     {
         var waveIn = new WaveInEvent
         {
@@ -30,6 +38,11 @@ public static AudioCaptureSession Start()
         };
 
         var session = new AudioCaptureSession(waveIn);
+        if (samplesAvailable is not null)
+        {
+            session.SamplesAvailable += samplesAvailable;
+        }
+
         try
         {
             waveIn.StartRecording();
@@ -62,13 +75,26 @@ public async Task<float[]> StopAsync(CancellationToken cancellationToken = defau
 
     private void OnDataAvailable(object? sender, WaveInEventArgs e)
     {
+        var samples = new float[e.BytesRecorded / 2];
+        for (var index = 0; index < samples.Length; index++)
+        {
+            var byteIndex = index * 2;
+            var sample = (short)(e.Buffer[byteIndex] | e.Buffer[byteIndex + 1] << 8);
+            samples[index] = sample / 32768f;
+        }
+
         lock (_gate)
         {
-            for (var index = 0; index + 1 < e.BytesRecorded; index += 2)
-            {
-                var sample = (short)(e.Buffer[index] | e.Buffer[index + 1] << 8);
-                _samples.Add(sample / 32768f);
-            }
+            _samples.AddRange(samples);
+        }
+
+        try
+        {
+            SamplesAvailable?.Invoke(this, new AudioSamplesAvailableEventArgs(samples));
+        }
+        catch
+        {
+            // Never let a UI/recognizer callback terminate the NAudio capture thread.
         }
     }
 
@@ -104,6 +130,7 @@ public void Dispose()
 
         _waveIn.DataAvailable -= OnDataAvailable;
         _waveIn.RecordingStopped -= OnRecordingStopped;
+        SamplesAvailable = null;
         _waveIn.Dispose();
         _stopped.TrySetResult();
     }
diff --git a/native-app/src/VoxType.NativeCore/ModelCatalog.cs b/native-app/src/VoxType.NativeCore/ModelCatalog.cs
index 9275f08..9b2230f 100644
--- a/native-app/src/VoxType.NativeCore/ModelCatalog.cs
+++ b/native-app/src/VoxType.NativeCore/ModelCatalog.cs
@@ -35,6 +35,17 @@ public ModelCatalogEntry GetDefaultSenseVoice()
                ?? throw new InvalidDataException("模型目录未配置可用的 SenseVoice 模型。");
     }
 
+    public ModelCatalogEntry GetDefaultStreamingModel()
+    {
+        return Models.FirstOrDefault(model =>
+                   model.Supported && model.Default
+                                   && model.Type.Equals("streaming-zipformer-ctc", StringComparison.OrdinalIgnoreCase))
+               ?? Models.FirstOrDefault(model =>
+                   model.Supported
+                   && model.Type.Equals("streaming-zipformer-ctc", StringComparison.OrdinalIgnoreCase))
+               ?? throw new InvalidDataException("模型目录未配置可用的中文流式模型。");
+    }
+
     private static readonly JsonSerializerOptions JsonOptions = new()
     {
         PropertyNameCaseInsensitive = true,
@@ -88,4 +99,7 @@ public sealed class ModelFileSpec
 
     [JsonPropertyName("required")]
     public bool Required { get; init; }
+
+    [JsonPropertyName("sha256")]
+    public string Sha256 { get; init; } = "";
 }
diff --git a/native-app/src/VoxType.NativeCore/ModelManager.cs b/native-app/src/VoxType.NativeCore/ModelManager.cs
index 68f2c68..fef0596 100644
--- a/native-app/src/VoxType.NativeCore/ModelManager.cs
+++ b/native-app/src/VoxType.NativeCore/ModelManager.cs
@@ -1,5 +1,6 @@
 using System.Net;
 using System.Net.Http.Headers;
+using System.Security.Cryptography;
 
 namespace VoxType.NativeCore;
 
@@ -29,7 +30,7 @@ public static bool IsInstalled(ModelCatalogEntry model)
                && requiredFiles.All(file =>
                {
                    ValidateFileName(file.Name);
-                   return IsComplete(Path.Combine(directory, file.Name), file.Size);
+                   return IsComplete(Path.Combine(directory, file.Name), file.Size, file.Sha256);
                });
     }
 
@@ -60,7 +61,7 @@ public async Task<string> EnsureInstalledAsync(
             cancellationToken.ThrowIfCancellationRequested();
             ValidateFileName(file.Name);
             var destination = Path.Combine(directory, file.Name);
-            if (IsComplete(destination, file.Size))
+            if (IsComplete(destination, file.Size, file.Sha256))
             {
                 completedBytes += file.Size ?? new FileInfo(destination).Length;
                 Report(progress, completedBytes, knownTotal, $"已存在：{file.Name}");
@@ -72,6 +73,7 @@ await DownloadFileAsync(
                     url,
                     destination,
                     file.Size,
+                    file.Sha256,
                     completedBytes,
                     knownTotal,
                     progress,
@@ -94,6 +96,7 @@ private async Task DownloadFileAsync(
         string url,
         string destination,
         long? expectedSize,
+        string expectedSha256,
         long completedBeforeFile,
         long knownTotal,
         IProgress<ModelDownloadProgress>? progress,
@@ -161,16 +164,24 @@ private async Task DownloadFileAsync(
         }
 
         await output.FlushAsync(cancellationToken).ConfigureAwait(false);
+        await output.DisposeAsync().ConfigureAwait(false);
+        await input.DisposeAsync().ConfigureAwait(false);
         if (expectedSize is > 0 && downloaded != expectedSize)
         {
             throw new InvalidDataException(
                 $"{Path.GetFileName(destination)} 下载不完整：{downloaded}/{expectedSize} bytes");
         }
 
+        if (!HasExpectedSha256(partPath, expectedSha256))
+        {
+            File.Delete(partPath);
+            throw new InvalidDataException($"{Path.GetFileName(destination)} SHA-256 校验失败。");
+        }
+
         File.Move(partPath, destination, true);
     }
 
-    private static bool IsComplete(string path, long? expectedSize)
+    private static bool IsComplete(string path, long? expectedSize, string expectedSha256)
     {
         if (!File.Exists(path))
         {
@@ -178,7 +189,20 @@ private static bool IsComplete(string path, long? expectedSize)
         }
 
         var length = new FileInfo(path).Length;
-        return expectedSize is > 0 ? length == expectedSize : length > 0;
+        var hasExpectedSize = expectedSize is > 0 ? length == expectedSize : length > 0;
+        return hasExpectedSize && HasExpectedSha256(path, expectedSha256);
+    }
+
+    private static bool HasExpectedSha256(string path, string expectedSha256)
+    {
+        if (string.IsNullOrWhiteSpace(expectedSha256))
+        {
+            return true;
+        }
+
+        using var input = File.OpenRead(path);
+        var actual = Convert.ToHexString(SHA256.HashData(input));
+        return actual.Equals(expectedSha256, StringComparison.OrdinalIgnoreCase);
     }
 
     private static void ValidateFileName(string fileName)
diff --git a/native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs b/native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs
new file mode 100644
index 0000000..9044ca3
--- /dev/null
+++ b/native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs
@@ -0,0 +1,337 @@
+using System.Diagnostics;
+using System.Threading.Channels;
+using SherpaOnnx;
+
+namespace VoxType.NativeCore;
+
+public sealed record StreamingRecognitionUpdate(
+    string Text,
+    bool IsEndpoint,
+    double AudioSeconds,
+    double ElapsedSeconds);
+
+public sealed record StreamingRecognitionResult(
+    string Text,
+    double AudioSeconds,
+    double ElapsedSeconds,
+    double? FirstTextSeconds);
+
+public sealed class StreamingVoiceEngine : IDisposable
+{
+    private readonly SemaphoreSlim _loadGate = new(1, 1);
+    private readonly object _sessionGate = new();
+    private OnlineRecognizer? _recognizer;
+    private string? _loadedModelDirectory;
+    private ChannelWriter<float[]>? _audioWriter;
+    private Task<StreamingRecognitionResult>? _sessionTask;
+    private bool _disposed;
+
+    public async Task WarmUpAsync(string modelDirectory, CancellationToken cancellationToken = default)
+    {
+        await _loadGate.WaitAsync(cancellationToken).ConfigureAwait(false);
+        try
+        {
+            ThrowIfDisposed();
+            lock (_sessionGate)
+            {
+                if (_sessionTask is not null)
+                {
+                    throw new InvalidOperationException("实时识别期间不能重新加载模型。");
+                }
+            }
+
+            await EnsureLoadedCoreAsync(modelDirectory, cancellationToken).ConfigureAwait(false);
+        }
+        finally
+        {
+            _loadGate.Release();
+        }
+    }
+
+    public void StartSession(Action<StreamingRecognitionUpdate>? onUpdate = null)
+    {
+        lock (_sessionGate)
+        {
+            ThrowIfDisposed();
+            if (_sessionTask is not null)
+            {
+                throw new InvalidOperationException("已有实时识别会话正在运行。");
+            }
+
+            var recognizer = _recognizer ?? throw new InvalidOperationException("实时识别模型尚未加载。");
+            var channel = Channel.CreateUnbounded<float[]>(new UnboundedChannelOptions
+            {
+                SingleReader = true,
+                SingleWriter = false,
+                AllowSynchronousContinuations = false,
+            });
+
+            _audioWriter = channel.Writer;
+            _sessionTask = Task.Run(() => ProcessSessionAsync(recognizer, channel.Reader, onUpdate));
+        }
+    }
+
+    public void AcceptAudio(float[] samples)
+    {
+        if (samples.Length == 0)
+        {
+            return;
+        }
+
+        ChannelWriter<float[]> writer;
+        lock (_sessionGate)
+        {
+            ThrowIfDisposed();
+            writer = _audioWriter ?? throw new InvalidOperationException("实时识别会话尚未开始。");
+        }
+
+        if (!writer.TryWrite(samples))
+        {
+            throw new InvalidOperationException("实时识别会话已经结束。");
+        }
+    }
+
+    public async Task<StreamingRecognitionResult> StopSessionAsync()
+    {
+        Task<StreamingRecognitionResult> task;
+        lock (_sessionGate)
+        {
+            ThrowIfDisposed();
+            task = _sessionTask ?? throw new InvalidOperationException("没有正在运行的实时识别会话。");
+            _audioWriter?.TryComplete();
+            _audioWriter = null;
+        }
+
+        try
+        {
+            return await task.ConfigureAwait(false);
+        }
+        finally
+        {
+            lock (_sessionGate)
+            {
+                if (ReferenceEquals(_sessionTask, task))
+                {
+                    _sessionTask = null;
+                }
+            }
+        }
+    }
+
+    private static async Task<StreamingRecognitionResult> ProcessSessionAsync(
+        OnlineRecognizer recognizer,
+        ChannelReader<float[]> reader,
+        Action<StreamingRecognitionUpdate>? onUpdate)
+    {
+        using var stream = recognizer.CreateStream();
+        var watch = Stopwatch.StartNew();
+        var committedText = "";
+        var lastPublishedText = "";
+        long acceptedSamples = 0;
+        double? firstTextSeconds = null;
+
+        await foreach (var samples in reader.ReadAllAsync().ConfigureAwait(false))
+        {
+            stream.AcceptWaveform(AudioCaptureSession.SampleRate, samples);
+            acceptedSamples += samples.Length;
+            DecodeReady(recognizer, stream);
+
+            var partialText = recognizer.GetResult(stream).Text.Trim();
+            var isEndpoint = recognizer.IsEndpoint(stream);
+            var combinedText = JoinSegments(committedText, partialText);
+
+            if (!string.IsNullOrEmpty(combinedText) && firstTextSeconds is null)
+            {
+                firstTextSeconds = watch.Elapsed.TotalSeconds;
+            }
+
+            if (isEndpoint)
+            {
+                committedText = combinedText;
+                recognizer.Reset(stream);
+            }
+
+            if (!string.Equals(combinedText, lastPublishedText, StringComparison.Ordinal) || isEndpoint)
+            {
+                lastPublishedText = combinedText;
+                Publish(onUpdate, new StreamingRecognitionUpdate(
+                    combinedText,
+                    isEndpoint,
+                    acceptedSamples / (double)AudioCaptureSession.SampleRate,
+                    watch.Elapsed.TotalSeconds));
+            }
+        }
+
+        // Streaming models need right context after the microphone stops. Silence flushes
+        // the last partial without adding user audio to the reported duration.
+        stream.AcceptWaveform(AudioCaptureSession.SampleRate, new float[AudioCaptureSession.SampleRate]);
+        stream.InputFinished();
+        DecodeReady(recognizer, stream);
+
+        var finalText = JoinSegments(committedText, recognizer.GetResult(stream).Text.Trim());
+        if (!string.IsNullOrEmpty(finalText) && firstTextSeconds is null)
+        {
+            firstTextSeconds = watch.Elapsed.TotalSeconds;
+        }
+
+        Publish(onUpdate, new StreamingRecognitionUpdate(
+            finalText,
+            true,
+            acceptedSamples / (double)AudioCaptureSession.SampleRate,
+            watch.Elapsed.TotalSeconds));
+
+        return new StreamingRecognitionResult(
+            finalText,
+            acceptedSamples / (double)AudioCaptureSession.SampleRate,
+            watch.Elapsed.TotalSeconds,
+            firstTextSeconds);
+    }
+
+    private static void DecodeReady(OnlineRecognizer recognizer, OnlineStream stream)
+    {
+        while (recognizer.IsReady(stream))
+        {
+            recognizer.Decode(stream);
+        }
+    }
+
+    private async Task EnsureLoadedCoreAsync(string modelDirectory, CancellationToken cancellationToken)
+    {
+        var fullDirectory = Path.GetFullPath(modelDirectory);
+        if (_recognizer is not null
+            && string.Equals(_loadedModelDirectory, fullDirectory, StringComparison.OrdinalIgnoreCase))
+        {
+            return;
+        }
+
+        var modelPath = Path.Combine(fullDirectory, "model.int8.onnx");
+        var tokensPath = Path.Combine(fullDirectory, "tokens.txt");
+        if (!File.Exists(modelPath) || !File.Exists(tokensPath))
+        {
+            throw new FileNotFoundException("中文流式 Zipformer 模型文件不完整。", fullDirectory);
+        }
+
+        var config = new OnlineRecognizerConfig();
+        config.FeatConfig.SampleRate = AudioCaptureSession.SampleRate;
+        config.FeatConfig.FeatureDim = 80;
+        config.ModelConfig.Zipformer2Ctc.Model = modelPath;
+        config.ModelConfig.Tokens = tokensPath;
+        config.ModelConfig.NumThreads = Math.Clamp(Environment.ProcessorCount / 2, 1, 4);
+        config.ModelConfig.Provider = "cpu";
+        config.ModelConfig.Debug = 0;
+        config.DecodingMethod = "greedy_search";
+        config.MaxActivePaths = 4;
+        config.EnableEndpoint = 1;
+        config.Rule1MinTrailingSilence = 2.4f;
+        config.Rule2MinTrailingSilence = 0.8f;
+        config.Rule3MinUtteranceLength = 20f;
+
+        var next = await Task.Run(() =>
+        {
+            var recognizer = new OnlineRecognizer(config);
+            using var stream = recognizer.CreateStream();
+            stream.AcceptWaveform(AudioCaptureSession.SampleRate, new float[AudioCaptureSession.SampleRate]);
+            stream.InputFinished();
+            DecodeReady(recognizer, stream);
+            return recognizer;
+        }, cancellationToken).ConfigureAwait(false);
+
+        OnlineRecognizer? previous;
+        lock (_sessionGate)
+        {
+            if (_disposed)
+            {
+                next.Dispose();
+                throw new ObjectDisposedException(nameof(StreamingVoiceEngine));
+            }
+
+            if (_sessionTask is not null)
+            {
+                next.Dispose();
+                throw new InvalidOperationException("实时识别期间不能替换模型。");
+            }
+
+            previous = _recognizer;
+            _recognizer = next;
+            _loadedModelDirectory = fullDirectory;
+        }
+
+        previous?.Dispose();
+    }
+
+    private static string JoinSegments(string committed, string current)
+    {
+        if (string.IsNullOrWhiteSpace(committed))
+        {
+            return current;
+        }
+
+        if (string.IsNullOrWhiteSpace(current))
+        {
+            return committed;
+        }
+
+        var needsSpace = char.IsAsciiLetterOrDigit(committed[^1]) && char.IsAsciiLetterOrDigit(current[0]);
+        return needsSpace ? $"{committed} {current}" : committed + current;
+    }
+
+    private static void Publish(
+        Action<StreamingRecognitionUpdate>? onUpdate,
+        StreamingRecognitionUpdate update)
+    {
+        try
+        {
+            onUpdate?.Invoke(update);
+        }
+        catch
+        {
+            // UI observers cannot be allowed to fault the native decoding worker.
+        }
+    }
+
+    private void ThrowIfDisposed()
+    {
+        ObjectDisposedException.ThrowIf(_disposed, this);
+    }
+
+    public void Dispose()
+    {
+        Task<StreamingRecognitionResult>? sessionTask;
+        lock (_sessionGate)
+        {
+            if (_disposed)
+            {
+                return;
+            }
+
+            _disposed = true;
+            _audioWriter?.TryComplete();
+            _audioWriter = null;
+            sessionTask = _sessionTask;
+        }
+
+        if (sessionTask is not null)
+        {
+            try
+            {
+                sessionTask.GetAwaiter().GetResult();
+            }
+            catch
+            {
+                // Window shutdown still needs to release the native recognizer.
+            }
+        }
+
+        _loadGate.Wait();
+        try
+        {
+            _recognizer?.Dispose();
+            _recognizer = null;
+        }
+        finally
+        {
+            _loadGate.Release();
+            _loadGate.Dispose();
+        }
+    }
+}
diff --git a/native-app/src/VoxType.NativeSmoke/Program.cs b/native-app/src/VoxType.NativeSmoke/Program.cs
index 1d1f13c..e327bb1 100644
--- a/native-app/src/VoxType.NativeSmoke/Program.cs
+++ b/native-app/src/VoxType.NativeSmoke/Program.cs
@@ -3,18 +3,19 @@
 
 if (args.Length == 0 || args.Contains("--help", StringComparer.OrdinalIgnoreCase))
 {
-    Console.WriteLine("Usage: VoxType.NativeSmoke <audio.wav> [model-directory] [repeat-count]");
-    Console.WriteLine($"Default model directory: {Path.Combine(VoxTypePaths.ModelsRoot, "sensevoice")}");
+    Console.WriteLine("Usage: VoxType.NativeSmoke <audio.wav> [model-directory] [repeat-count] [--realtime]");
+    Console.WriteLine($"Default model directory: {Path.Combine(VoxTypePaths.ModelsRoot, "zipformer-zh-streaming-small")}");
     return 0;
 }
 
 var audioPath = Path.GetFullPath(args[0]);
 var modelDirectory = args.Length > 1
     ? Path.GetFullPath(args[1])
-    : Path.Combine(VoxTypePaths.ModelsRoot, "sensevoice");
+    : Path.Combine(VoxTypePaths.ModelsRoot, "zipformer-zh-streaming-small");
 var repeatCount = args.Length > 2 && int.TryParse(args[2], out var parsedRepeatCount)
     ? Math.Clamp(parsedRepeatCount, 1, 20)
     : 1;
+var realtimePlayback = args.Length > 3 && args[3].Equals("--realtime", StringComparison.OrdinalIgnoreCase);
 
 if (!File.Exists(audioPath))
 {
@@ -29,7 +30,7 @@
     Console.WriteLine($"Audio: {audioSeconds:F2}s, samples={samples.Length}");
     Console.WriteLine($"Model: {modelDirectory}");
 
-    using var engine = new SenseVoiceEngine();
+    using var engine = new StreamingVoiceEngine();
     var loadWatch = Stopwatch.StartNew();
     await engine.WarmUpAsync(modelDirectory);
     loadWatch.Stop();
@@ -39,12 +40,41 @@
     for (var index = 1; index <= repeatCount; index++)
     {
         var watch = Stopwatch.StartNew();
-        var text = await engine.TranscribeAsync(modelDirectory, samples);
+        var latestText = "";
+        var firstPartial = "";
+        var updateCount = 0;
+        engine.StartSession(update =>
+        {
+            latestText = update.Text;
+            if (!string.IsNullOrWhiteSpace(update.Text))
+            {
+                Interlocked.CompareExchange(ref firstPartial, update.Text, "");
+            }
+
+            Interlocked.Increment(ref updateCount);
+        });
+        const int chunkSize = AudioCaptureSession.SampleRate * 40 / 1000;
+        for (var offset = 0; offset < samples.Length; offset += chunkSize)
+        {
+            var count = Math.Min(chunkSize, samples.Length - offset);
+            var chunk = new float[count];
+            Array.Copy(samples, offset, chunk, 0, count);
+            engine.AcceptAudio(chunk);
+            if (realtimePlayback)
+            {
+                await Task.Delay(40);
+            }
+        }
+
+        var result = await engine.StopSessionAsync();
         watch.Stop();
 
         Console.WriteLine($"Run {index}: {watch.Elapsed.TotalSeconds:F2}s, RTF={watch.Elapsed.TotalSeconds / audioSeconds:F3}");
-        Console.WriteLine($"Text: {text}");
-        hasText |= !string.IsNullOrWhiteSpace(text);
+        Console.WriteLine($"Text: {result.Text}");
+        Console.WriteLine($"Updates: {updateCount}, first text: {result.FirstTextSeconds?.ToString("F2") ?? "n/a"}s");
+        Console.WriteLine($"First partial: {firstPartial}");
+        Console.WriteLine($"Last partial: {latestText}");
+        hasText |= !string.IsNullOrWhiteSpace(result.Text);
     }
 
     return hasText ? 0 : 3;

From 35066fed04d4eba972278bd8d9131f5d3ba1b354 Mon Sep 17 00:00:00 2001
From: Cea <61349137+ceastld@users.noreply.github.com>
Date: Fri, 7 Aug 2026 21:28:05 +0800
Subject: [PATCH 4/4] fix: preserve streaming recognition context

---
 native-app/BENCHMARKS.md                      |  3 +-
 native-app/README.md                          |  2 +-
 .../StreamingVoiceEngine.cs                   | 53 ++++++++++++++++++-
 3 files changed, 54 insertions(+), 4 deletions(-)

diff --git a/native-app/BENCHMARKS.md b/native-app/BENCHMARKS.md
index f858f96..62910ce 100644
--- a/native-app/BENCHMARKS.md
+++ b/native-app/BENCHMARKS.md
@@ -17,7 +17,7 @@
 |------|------|
 | 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 | 8 次文字更新，首个 partial 约 0.91 s |
+| 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 |
@@ -28,6 +28,7 @@
 ## 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。
diff --git a/native-app/README.md b/native-app/README.md
index 78cb00c..372c767 100644
--- a/native-app/README.md
+++ b/native-app/README.md
@@ -47,5 +47,5 @@ dotnet run --project ./src/VoxType.NativeSmoke -c Release -- <audio.wav> [model-
 - 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。
-- 端点检测在连续静音后确认当前分段并继续同一录音会话；点击停止时额外送入右侧静音上下文以刷新最终结果。
+- 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。
diff --git a/native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs b/native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs
index 9044ca3..90f1d96 100644
--- a/native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs
+++ b/native-app/src/VoxType.NativeCore/StreamingVoiceEngine.cs
@@ -130,6 +130,13 @@ private static async Task<StreamingRecognitionResult> ProcessSessionAsync(
         long acceptedSamples = 0;
         double? firstTextSeconds = null;
 
+        // Match sherpa-onnx's reference streaming decoder: give the first spoken
+        // frame left context so an immediate utterance is not treated as a clipped start.
+        stream.AcceptWaveform(
+            AudioCaptureSession.SampleRate,
+            new float[AudioCaptureSession.SampleRate * 3 / 10]);
+        DecodeReady(recognizer, stream);
+
         await foreach (var samples in reader.ReadAllAsync().ConfigureAwait(false))
         {
             stream.AcceptWaveform(AudioCaptureSession.SampleRate, samples);
@@ -164,7 +171,9 @@ private static async Task<StreamingRecognitionResult> ProcessSessionAsync(
 
         // Streaming models need right context after the microphone stops. Silence flushes
         // the last partial without adding user audio to the reported duration.
-        stream.AcceptWaveform(AudioCaptureSession.SampleRate, new float[AudioCaptureSession.SampleRate]);
+        stream.AcceptWaveform(
+            AudioCaptureSession.SampleRate,
+            new float[AudioCaptureSession.SampleRate * 6 / 10]);
         stream.InputFinished();
         DecodeReady(recognizer, stream);
 
@@ -223,7 +232,7 @@ private async Task EnsureLoadedCoreAsync(string modelDirectory, CancellationToke
         config.MaxActivePaths = 4;
         config.EnableEndpoint = 1;
         config.Rule1MinTrailingSilence = 2.4f;
-        config.Rule2MinTrailingSilence = 0.8f;
+        config.Rule2MinTrailingSilence = 1.2f;
         config.Rule3MinUtteranceLength = 20f;
 
         var next = await Task.Run(() =>
@@ -261,6 +270,8 @@ private async Task EnsureLoadedCoreAsync(string modelDirectory, CancellationToke
 
     private static string JoinSegments(string committed, string current)
     {
+        committed = committed.TrimEnd();
+        current = current.TrimStart();
         if (string.IsNullOrWhiteSpace(committed))
         {
             return current;
@@ -271,10 +282,48 @@ private static string JoinSegments(string committed, string current)
             return committed;
         }
 
+        // An online model can repeat a few words after endpoint reset because its
+        // acoustic right context overlaps the next stream. Remove only a reasonably
+        // long suffix/prefix match; short repetitions can be intentional speech.
+        var maximumOverlap = Math.Min(24, Math.Min(committed.Length, current.Length));
+        for (var length = maximumOverlap; length >= 6; length--)
+        {
+            var left = committed.AsSpan(committed.Length - length, length);
+            var right = current.AsSpan(0, length);
+            var allowedDifferences = Math.Clamp(length / 8, 1, 2);
+            if (CountDifferences(left, right, allowedDifferences) <= allowedDifferences)
+            {
+                return committed + current[length..];
+            }
+        }
+
         var needsSpace = char.IsAsciiLetterOrDigit(committed[^1]) && char.IsAsciiLetterOrDigit(current[0]);
         return needsSpace ? $"{committed} {current}" : committed + current;
     }
 
+    private static int CountDifferences(
+        ReadOnlySpan<char> left,
+        ReadOnlySpan<char> right,
+        int stopAfter)
+    {
+        var differences = 0;
+        for (var index = 0; index < left.Length; index++)
+        {
+            if (left[index] == right[index])
+            {
+                continue;
+            }
+
+            differences++;
+            if (differences > stopAfter)
+            {
+                break;
+            }
+        }
+
+        return differences;
+    }
+
     private static void Publish(
         Action<StreamingRecognitionUpdate>? onUpdate,
         StreamingRecognitionUpdate update)