XiLuoLin

English | 简体中文

XiLuoLin is an open-source AI voice input assistant for office work, writing, and programming. It turns short voice recordings into ready-to-use text and reduces the effort spent typing, editing, and polishing through persona-based rewriting, hotwords, local history, and desktop text delivery. An application recording is limited to 28 seconds; externally selected audio is limited to 30 seconds.

XiLuoLin is independently initiated and actively maintained. Community participation through Issues, Discussions, and Pull Requests is welcome.

Stable Release Downloads

v0.2.0 is the current stable release:

• macOS 13+ on Apple Silicon: an ad-hoc-signed, non-notarized DMG that requires manual approval on first launch.
• Windows 10/11 x64: an unsigned NSIS installer that may trigger Microsoft Defender SmartScreen.
• These packages may still trigger operating-system warnings because macOS is not notarized and Windows is not code-signed.
• Intel Macs, Windows ARM64, Linux, app stores, and in-app updates are out of scope for this release.

Download only from GitHub Releases and verify the file against SHA256SUMS.txt from the same release. See the macOS installation guide, Windows installation guide, and changelog before installing.

Product Direction

XiLuoLin focuses on the complete workflow from speaking an idea to getting text that can be used immediately:

• Voice capture: microphone recording, short-audio processing, global shortcuts, and recording status feedback; macOS can explicitly enable hold-Fn recording with short-tap cancellation.
• Speech recognition: Zhipu GLM-ASR-2512, OpenAI-compatible, Qwen-Audio, Qwen3-ASR, and an offline local Whisper model.
• Experimental live captions: disabled by default; an explicitly downloaded bilingual Zipformer mixed-quantization candidate (about 199.3 MB) provides incremental overlay text without replacing final ASR. Preview failures do not affect final recognition, history, or delivery.
• Persona-based rewriting: use polished built-in or custom personas, or select Verbatim Dictation to preserve the raw ASR wording while only normalizing whitespace.
• Hotword dictionary: enabled hotwords globally bias ASR; Zhipu and Qwen-Audio receive up to 100 native hotwords, Qwen3-ASR uses a system glossary, and OpenAI/local Whisper use soft prompts.
• Desktop delivery: clipboard and automatic paste output; when automatic paste fails, the result is copied to the clipboard and the overlay tells the user to paste manually.
• Local data: store history, personas, hotwords, settings, and personal productivity statistics locally.
• Open extensibility: maintain clear Provider and business-module boundaries for additional cloud services, local models, and cross-platform integrations.

Current Status

XiLuoLin is under active development. The core modules are available, while reliability, cross-platform verification, usability, packaging, and release workflows continue to evolve.

Major capabilities already implemented:

• Tauri v2, React, and TypeScript desktop application foundation
• SQLite local data layer and operating-system credential storage
• Built-in personas, custom personas, and default persona selection
• Hotword dictionary, input history, and statistics
• Zhipu, OpenAI-compatible, Qwen-Audio, Qwen3-ASR, and local Whisper ASR Providers
• Zhipu, OpenAI-compatible, and Qwen text-rewriting Providers
• Recording, global shortcuts, a recording indicator, and a 25-second warning / 28-second automatic stop for application recordings
• An experimental React live-transcript overlay with explicit, checksummed local model downloads; the candidate model remains No-Go for production redistribution until its training-data license chain is auditable
• Clipboard delivery, automatic paste, and error feedback
• Home, persona, hotword, and settings pages
• TypeScript, frontend build, Rust formatting, compilation, and test checks

Current priorities:

• Verify microphone, shortcut, credential-store, and cross-application delivery behavior across operating systems
• Improve the home-page voice entry point and recording-state experience; the global shortcut remains the primary input entry today
• Validate the stable installers and continue improving release and compatibility documentation
• Improve Provider configuration, failure recovery, and automated testing
• Continue improving contributor documentation, Issue management, and technical decision records

Documentation

The detailed product and engineering documents are currently maintained in Chinese:

• Documentation index
• Product requirements
• Technical design
• Usage and verification guide
• Troubleshooting guide
• ASR quality evaluation and desktop acceptance
• macOS Apple Silicon build and installation
• Windows x64 build and installation
• Changelog
• Project roadmap
• Contribution guide
• Security policy
• Code of Conduct

Technology Stack

• Desktop framework: Tauri v2
• Frontend: React 19, TypeScript, and Vite
• UI: Tailwind CSS, shadcn/ui, and Radix UI
• Local storage: SQLite, Tauri Store, and the operating-system credential store
• Audio: cpal and hound; statically linked Apache-2.0 sherpa-onnx for optional local live captions
• External services: configurable ASR and text-processing Providers

Requirements

• Node.js 20+
• pnpm 10+
• Rust stable toolchain
• Windows: Microsoft Visual Studio C++ Build Tools and WebView2 Runtime
• macOS 13+: microphone permission; cross-app automatic paste also requires Accessibility permission
• Windows: microphone permission; cross-integrity-level input can be restricted by the system

Local Development

git clone https://github.com/qinyu765/xiluolin.git
cd xiluolin
pnpm install --frozen-lockfile
pnpm check
pnpm tauri dev

Common commands:

Command	Purpose
pnpm dev	Start the frontend development server
pnpm typecheck	Run TypeScript type checking
pnpm build	Type-check and build the frontend
pnpm check:rust	Run Rust formatting, compilation, and tests
pnpm check	Run the complete frontend and Rust quality checks
pnpm tauri dev	Start the desktop application in development mode
pnpm release:check	Verify frontend, Cargo, Tauri, and optional release-tag versions
pnpm eval:asr	Compute CER, hotword recall, punctuation F1, and latency for a private benchmark
pnpm tauri:build:macos:arm64	Build the macOS 13+ Apple Silicon app and DMG
pnpm tauri:build:windows:x64	Build the Windows 10/11 x64 NSIS installer on Windows

GitHub Actions runs frontend, Windows/macOS Rust, dependency-security, and secret checks for pushes to main and Pull Requests targeting main; release Pull Requests also build the macOS DMG and Windows NSIS installer. Changes involving recording, global shortcuts, credentials, or text delivery still require manual verification in a desktop environment.

Configuration and Usage

1. Start the application and open Settings.
2. Configure Zhipu GLM-ASR-2512, OpenAI-compatible, Qwen-Audio, Qwen3-ASR, or the offline local Whisper model.
3. Configure Zhipu, OpenAI-compatible, or Qwen text processing.
4. Optionally download the verified bilingual Zipformer model under Settings → Model configuration to enable live overlay captions.
5. Select the microphone, shortcuts, and output method. On macOS, hold-Fn recording can be enabled after Accessibility permission is granted.
6. Select Verbatim Dictation when no text-model rewriting is wanted, or use a polished built-in/custom persona.
7. Add project names, personal names, and technical terms that require more accurate recognition.
8. Place the cursor in a target application and use a global shortcut to start voice input.

Settings are saved automatically: switches, selects, and shortcuts save immediately; text and API key fields save about 600 ms after typing stops and flush on blur. If a save fails, retry it from the Settings status indicator.

See the usage and verification guide for detailed setup, validation paths, and failure scenarios.

Privacy and Security

• Audio is sent only to the ASR Provider explicitly configured by the user.
• Raw recognized text is sent only to the text-processing Provider explicitly configured by the user.
• API keys are stored in Windows Credential Manager, macOS Keychain, or another operating-system-native credential store.
• History, personas, hotwords, and statistics are stored in local SQLite by default and are not uploaded to a XiLuoLin server.
• Temporary recordings created by the application are removed after either successful or failed processing. User-selected external audio files are never deleted by this cleanup logic.
• Logs must not contain API keys, complete user text, or complete recording paths.
• Live captions run locally. Preview text is never used as a final-result fallback and is not written to logs.

Before using a third-party Provider, review its privacy policy, data-retention rules, and terms of service. Report security concerns according to SECURITY.md.

Contributing

Contributions are welcome in many forms:

• Bug reports, reproducible cases, and cross-platform compatibility feedback
• Product proposals, interaction improvements, and accessibility work
• Provider, recording, shortcut, output, and local-storage improvements
• Tests, documentation, translations, and release-process improvements

Before contributing, read CONTRIBUTING.md and CODE_OF_CONDUCT.md. For larger features or architectural changes, open an Issue first, then submit a focused Pull Request from a short-lived branch to main.

When a change affects project positioning, capabilities, setup, usage, privacy, compatibility, or contribution instructions, update both the English README.md and Chinese README.zh-CN.md in the same Pull Request.

Project Governance

• main is the stable development baseline, and changes are merged through branches and Pull Requests.
• The roadmap communicates direction and does not guarantee fixed delivery dates.
• Proposals are evaluated according to user value, reliability, privacy risk, maintenance cost, and architectural consistency.
• Historical development records are kept under docs/dev/. References to competitions, demos, MVPs, or older workflows describe their original context and do not define the project's current direction.

License

XiLuoLin is open source under the MIT License.