Zephyr

Zephyr is a Windows AI voice input assistant built with Tauri 2, Rust, Preact, and TypeScript.

It is not a Windows TSF input method. It is a lightweight helper tool: hold a global shortcut, speak, watch a floating preview, and release to deliver the final cloud ASR transcript into the active app.

Zephyr focuses on a fast preview-first voice input flow. The core input path stays native/Rust-side, while the Tauri front end provides settings, history, hotword management, and visual feedback.

Features

• Hold-to-talk global shortcut, default Ctrl+Alt+Space.
• Volcengine bidirectional streaming ASR over WebSocket, default bigmodel_async.
• Preview Mode: partial text appears in a non-focus-stealing floating preinput window.
• Final text is delivered through Unicode SendInput by default, without modifying the clipboard.
• Dynamic shortcut settings with conflict detection.
• Recognition behavior switches for punctuation, ITN, semantic smoothing, and first-character acceleration.
• Local history with search, edit, copy, delete, and clear.
• Hotword and context management with optional DeepSeek-based organization.
• Light ASCII visual settings UI with tray integration.

Platform

The current target is Windows only.

Tested development assumptions:

• Windows 10/11
• Rust MSVC toolchain
• Visual Studio Build Tools or Visual Studio Community with C++ desktop workload
• Node.js and npm

Quick Start

git clone https://github.com/zhongwater123/Zephyr.git
cd Zephyr
npm install
npm run tauri dev

For Rust-only checks:

cd src-tauri
cargo test

For production build:

npm run package:windows

The command runs the release gates, builds a current-user NSIS installer, and writes a SHA-256 release manifest next to it. See Windows test installer and upgrades.

Configuration

Open the settings drawer in the app and configure the recognition service.

Default ASR settings:

• Endpoint: wss://openspeech.bytedance.com/api/v3/sauc/bigmodel_async
• Resource ID: volc.bigasr.sauc.duration
• Model: bigmodel
• Language: zh-CN
• Auth mode: old console App Key + Access Key by default; API Key mode is also supported.

Secrets are stored in the operating system keyring. They are not written to config.json.

See docs/configuration.md and docs/provider-volcengine.md.

Data And Privacy

Zephyr handles voice audio and recognized text. Some data is sent to external cloud services when the related feature is enabled:

• Audio is streamed to the configured ASR provider during a voice session.
• Final transcripts can be stored locally in history.db when history is enabled.
• History text can be sent to DeepSeek only when the hotword agent is enabled or manually triggered.

See PRIVACY.md before using real personal or sensitive data.

Documentation

• Architecture
• Development Setup
• Configuration
• Volcengine ASR Provider
• Hotwords And Context
• Troubleshooting
• Contributing
• Security

Contributing

Issues, bug reports, usability notes, and small focused pull requests are welcome.

Good first areas to help with:

• Windows compatibility and edge-case testing across common apps.
• Volcengine ASR protocol robustness.
• Preview overlay latency and positioning.
• History, hotword, and context-management UX.
• Documentation and setup improvements.

Please avoid committing API keys, personal transcripts, local databases, generated build output, or dependency folders.

Repository Hygiene

Generated and local-only files should not be committed:

• node_modules/
• dist/
• target/
• target-check/
• .vs/
• sauc_go/
• sauc_go.zip
• .env*
• logs and temporary files

License

MIT. See LICENSE.