BaiYunGe

Project Overview

BaiYunGe is an offline voice input software for Windows 10/11 x64. Press the hotkey, speak, and release to type text. Recognition runs entirely on-device, with audio and text never leaving your machine.

Key Features

• Offline recognition: Local Qwen3-ASR-1.7B model + built-in llama.cpp inference (Vulkan GPU acceleration, automatically prefers discrete GPU on multi-GPU machines, falls back to CPU without one)
• Two hotkey modes: Hold to talk and release to recognize (hold, default) / press to start and press again to stop (toggle); default hotkey Ctrl+Win, customizable recording
• Direct typing: Recognized text is typed into the current focus input field via SendInput; automatically copied to clipboard and pasted when focus is not editable
• Dictionary hotwords: Custom "alias → target word" mapping, automatically replacing recognition results
• Bilingual Chinese-English interface + day/night themes
• One-click model download in-app: Resumable downloads, SHA-256 verification, auto-retry 3 times on failure
• In-app update check: One-click download, silent overwrite install, retaining all configuration

Target Audience

• Users who need efficient voice input
• Privacy-conscious users who don't want audio uploaded to the cloud
• Windows 10/11 x64 users

Quick Start

1. Download 白云歌_BaiYunGe_x.x.x_x64_Setup.exe from Releases
2. Double-click to install (Chinese wizard, optional install directory)
3. First run automatically enters the "Local Model" page; click "Install Model" to download the speech model (~2.4 GB)
4. The software then silently resides in the tray; hold Ctrl+Win to speak

Core Advantages

• Fully offline; audio and text never leave your machine
• Vulkan GPU acceleration for fast recognition
• Dictionary hotwords improve proper noun recognition accuracy
• Auto-start on boot, silently resides in the tray

System Requirements

• OS: Windows 10/11 x64
• Additional dependencies: .NET 8 SDK (for building from source); ~2.4 GB speech model download required on first run