VoiceInput

v0.1.4-preview · Windows Local Voice Input

Application version: 0.1.4. This release focuses on usable, repeatable recording and a reproducible distribution path. It keeps the -preview label so hardware-specific feedback can continue. See the documentation index.

English | 简体中文

Latest Release | Chinese release notes | English release notes | Fix and validation report

⚠️Preview Release — This is an early preview version for testing and feedback. Expect bugs and breaking changes before the stable release.

A privacy-first, fully-local voice input tool for Windows. Speak into your microphone and your words appear wherever your cursor is — no typing needed. Speech recognition runs entirely on your GPU via Qwen3-ASR, so no audio ever leaves your computer.

This is a Windows desktop preview built with React, Tauri, Rust, and Python. The desktop layer handles recording, hotkeys, and safe text entry; the Python service listens only on a random local loopback port and runs the model on a local NVIDIA GPU.

Current status

✅ Source version is synchronized at 0.1.4 across React, Rust/Tauri, and the Python sidecar
✅ Local NVIDIA GPU, Qwen3-ASR loading, microphone recording, and automatic text entry have been verified
✅ WASAPI capture is owned by a dedicated worker and audio-level events no longer block the realtime callback
✅ The release flow validates fresh extraction, sidecar health, and a SHA-256 manifest
⚠️ Preview release: elevated windows, games, and custom controls may still reject simulated input

What is VoiceInput? (For Everyone)

VoiceInput is like having a friend who types everything you say. Instead of using your keyboard, you just:

Hold a key (Alt+V) on your keyboard
Talk into your microphone
Let go of the key
Your words magically appear where your cursor is blinking!

It works in most standard text fields — Word, chat boxes, and browser search bars. Some elevated windows, games, and custom-drawn controls may block simulated input and need separate validation.

Features

🎙️ Push-to-talk — Hold Alt+V, speak, release to transcribe and paste automatically
🖱️ Click to talk — Click the microphone button, click again to stop
🌐 Language switching — Alt+L cycles Auto / Chinese / English
🔒 100% offline — All recognition runs on the local GPU, zero cloud calls
📝 Custom terms — Define ASR misrecognition → correct-text mappings
🩺 Auto-recovery — Backend crash auto-restart + file logging
🖥️ System tray — Quick access to settings, logs, and quit
⚡ VAD auto-stop — Detects silence and stops recording automatically
📋 Clipboard-safe input — Returns to the prior window and types without altering clipboard data
🧪 Audio diagnostics — Includes a developer probe to distinguish microphone/input and Bluetooth speaker/output problems

System Requirements

Requirement	Detail
OS	Windows 10 1903+ / Windows 11
GPU	NVIDIA graphics card with CUDA support, ≥ 4 GB VRAM
Disk	~2.7 GB runtime + ~1.9 GB model (downloaded once); keep 6 GB free
RAM	8 GB+
Input	A microphone (built-in or external)

⚠️Important: An NVIDIA graphics card is required. Intel/AMD integrated graphics are not supported. If you're not sure what GPU you have, look for "NVIDIA" in your Task Manager → Performance tab.

Quick Start (Step by Step)

Step 1: Download

Go to the Releases page
This release currently has no downloadable runtime asset because the complete 0.1.4 onedir ZIP is about 2.73 GB and exceeds the GitHub Release per-asset limit. GitHub's source archives are not runnable packages; build from source using the instructions below or use a complete ZIP hosted in an external large-file store.
If the release includes VoiceInput-v0.1.4-preview-win64.zip, download it. If the asset is temporarily unavailable, build from source using the instructions below. GitHub's generated source archive is not runnable by itself.

Step 2: Extract

Find the downloaded .zip file (usually in your Downloads folder)
Right-click the zip file → Extract All...
Choose a location, for example: C:\VoiceInput
Click Extract

💡 Tip: Don't put it in a folder that needs admin permission (like C:\Program Files\), or it might not work properly.

Step 3: Run

Open the extracted folder
Double-clickvoiceinput.exe
A small floating window will appear on your screen

🛡️ If Windows shows a "Windows protected your PC" warning, click More info → Run anyway. This is normal for apps that aren't from the Microsoft Store.

Step 4: Download the Model (First Launch Only)

The first time you run VoiceInput, it needs to download the AI model (about 1.9 GB). This happens only once.

The app will show a download screen
Choose ModelScope (recommended, faster for China) or HuggingFace (international)
Click 开始下载 (Start Download)
Wait for the download to finish (this depends on your internet speed, usually 5-20 minutes)
Click 加载模型 (Load Model) — this takes about 10-20 seconds

Step 5: Start Talking!

Open any program where you can type (e.g., a Word document, a chat box, a search bar)
Click where you want the text to appear
Press and hold Alt+V on your keyboard
Speak into your microphone
Release Alt+V
Your words will appear automatically!

How to Use

The Floating Window

The small window on your screen has these parts:

Microphone button (big circle): Click to start/stop recording
Language badge (top-left, shows "Auto"/"中"/"EN"): Click to switch language
Gear icon (top-right): Click to open Settings
X icon (top-right): Click to hide the window (app keeps running in system tray)

Keyboard Shortcuts

Shortcut	What it does
`Alt+V` (hold)	Push-to-talk: hold while speaking, release to transcribe
`Alt+L`	Switch language: Auto → Chinese → English → Auto

System Tray

A green microphone icon appears in your system tray (bottom-right of screen, near the clock). Right-click it to:

Show/hide the panel
Open Settings
Switch language
Load/unload the model (free GPU memory)
View log files
Quit the app

Settings

Open Settings by clicking the gear icon on the floating window, or right-click the tray icon → 设置.

Microphone Tab

Choose which microphone to use. Click Test to record a short sample and see if it works.

Hotkey Tab

Change the push-to-talk hotkey (default: Alt+V) and language-switch hotkey (default: Alt+L).

Audio Tab

Adjust sample rate, audio normalization, silence trimming, and silence threshold.

Advanced Tab

Model strategy: How the AI model uses GPU memory
- Balanced (default): Unloads after 30 minutes idle
- Performance: Always loaded (fastest response)
- Memory saver: Unloads after each use (lowest VRAM)
- Accurate: Maximum precision
Punctuation mode: How punctuation is handled
- Simple (default): Auto-adds periods at the end
- Raw: No changes
- Input method: Removes end punctuation (good for search boxes)
Auto space (Chinese-English): Adds spaces between Chinese and English
VAD silence detection: Automatically stops recording after 2 seconds of silence
Auto-start on boot: Start VoiceInput when Windows starts
Input delay, max recording time, request timeout (automatic input does not touch the clipboard)

Terms Tab

Add custom word corrections. For example, if the AI always mishears your name "Xiaoming" as "Xiao Ming", add a rule to fix it.

Known boundaries

Windows 10 1903+ / Windows 11 and an NVIDIA CUDA GPU are required; CPU, AMD, and Intel GPU inference are not supported.
The model download is about 1.9 GB; the release ZIP is about 2.7 GB, so plan for both separately.
Automatic entry depends on the target application accepting Windows input events; elevated windows, custom controls, and some games may reject it.
Compiled binaries and model caches are not committed to the repository. Download the complete release package and keep asr_backend/_internal beside the backend executable.
audio_hardware_probe is a developer diagnostic binary, not the normal application entry point; it reports the default devices and measured capture energy.

Troubleshooting

The app won't start / closes immediately

Most likely cause: No NVIDIA GPU. VoiceInput requires an NVIDIA graphics card. Check Task Manager → Performance tab for "NVIDIA".
Solution: Install the latest NVIDIA driver. If you have a laptop with both Intel and NVIDIA, set VoiceInput to use NVIDIA in NVIDIA Control Panel.

A dialog says "another instance is already running"

VoiceInput is already open. Look for the green microphone icon in your system tray (bottom-right of screen).

Recording starts but no text appears

Check: Is the model loaded? Open Settings → Advanced → Model strategy. If it says "not loaded", click the tray icon → 预加载模型.
Check: Is your microphone working? Open Settings → Microphone → Test.

Text appears in the wrong place

VoiceInput pastes text at your cursor position. Click where you want the text before speaking.
Some apps (like certain games) may block paste. Try clicking in a text box first.

Recognition is slow

First recognition after loading takes longer (the model is warming up). Subsequent ones are faster.
Try changing Model Strategy to Performance (always loaded) in Settings.
Make sure no other heavy GPU programs are running.

Download fails / is very slow

Choose ModelScope if you're in China (faster), or HuggingFace for international.
The download supports resume — if it stops, just click download again to continue.

Building from Source

Prerequisites

Node.js 20.19+ and npm
Rust (stable toolchain)
Python 3.10+ with backend dependencies (torch, fastapi, qwen_asr, ...)
Tauri CLI 2.x (npm install -D @tauri-apps/cli)
NVIDIA CUDA Toolkit 11.8+

Steps

# 1. Install frontend dependencies
npm install
# 2. Install backend dependencies and a fixed packager in a dedicated environment
python -m pip install -r backend\requirements.txt
python -m pip install pyinstaller==6.22.1
# 3. Build the Python ASR backend (PyInstaller onedir → src-tauri/binaries/asr_backend/)
.\build_backend.bat
# 4. Build the Tauri executable
npm run tauri -- build
# 5. Build and fresh-extraction validate the release zip
powershell -ExecutionPolicy Bypass -File .\build_release_zip.ps1

The release zip is output to .\release\VoiceInput-v0.1.4-preview-win64.zip.

Packaging note — The backend is an onedir runtime (~2.7 GB including torch and transformers) to avoid unpacking a giant executable on every launch. The project ships a ZIP; keep the complete asr_backend/_internal directory beside asr_backend.exe.

Architecture

voiceinput.exe (Tauri/Rust) ──spawn──▶ asr_backend.exe (Python/FastAPI)
│ │
├─ cpal (audio capture) ├─ Qwen3-ASR-0.6B (GPU inference)
├─ Windows SendInput (safe text entry) ├─ uvicorn (random 127.0.0.1 port)
├─ rdev (global hotkeys) └─ postprocess (term correction)
└─ React 18 (floating UI)

Layer	Tech	Key Files
Frontend	React 18 + TypeScript + Tailwind	`src/`
Native shell	Rust + Tauri 2	`src-tauri/src/`
ASR backend	Python + FastAPI + PyTorch	`backend/`

Configuration

User config lives at %LOCALAPPDATA%\VoiceInput\config.json. Logs at %LOCALAPPDATA%\VoiceInput\logs\.

Project Structure

voice-input-v2/
├─ src/ # React frontend
├─ src-tauri/ # Rust + Tauri native shell
│ ├─ src/ # Rust source
│ ├─ binaries/ # PyInstaller sidecar output (gitignored)
│ └─ tauri.conf.json # Tauri bundle config
├─ backend/ # Python ASR backend
│ ├─ server.py # FastAPI server
│ ├─ model_manager.py # Model download/load
│ ├─ postprocess.py # Text post-processing
│ └─ asr_backend.spec # PyInstaller spec
├─ resources/ # Default config, icon
├─ build_backend.bat # Build sidecar EXE
├─ build_release_zip.ps1 # Build release zip
└─ docs/PRD.md # Product requirements

License

MIT — see LICENSE.

Acknowledgements

Qwen3-ASR — the ASR model
Tauri — the app framework
PyInstaller — Python bundling

Changelog

v0.1.4-preview (Latest)

This release is focused on real usability, repeated use, and a reproducible release process. The desktop shell, recorder, Python sidecar, build scripts, and bilingual documentation are synchronized at 0.1.4.

Code and reliability:

🎙️ Dedicated WASAPI recorder thread — Each session creates, owns, and releases the cpal stream on one dedicated thread, reducing lifecycle conflicts between Tauri commands and Windows audio sessions.
📈 Decoupled audio-level delivery — The realtime callback only converts samples, maintains the bounded buffer, and stores an atomic level snapshot. A lightweight 50 ms sampler sends audio-level events so the waveform and microphone test remain responsive.
⏱️ VAD timing fix — Silence timing starts after the audio stream is actually running; WASAPI initialization time is not counted as user silence. Minimum-duration and continuous-silence guards remain enabled.
🩺 Better observability — Stop logs include total RMS/peak; VAD logs include current/maximum RMS. A separate hardware probe can play Bluetooth-speaker audio while measuring microphone capture.
🛡️ Startup cleanup — Recorder startup timeouts and failures wait for the worker to exit instead of leaving an orphan audio thread.

Release and documentation:

package.json, Cargo, Tauri, FastAPI, and sidecar startup logging are synchronized to 0.1.4.
The onedir release flow validates complete extraction, the relative asr_backend/_internal layout, the manifest, and backend health.
Added Chinese and English 0.1.4 release notes, upgrade guides, and validation records; 0.1.3 and 0.1.2 notes remain historical.

Upgrade notes:

Replace voiceinput.exe and the complete asr_backend directory together; do not replace only one component.
Existing model caches remain compatible and do not need to be downloaded again. Re-select or re-download only if the model directory is damaged.
After upgrading, use Settings → Microphone → Test once, then complete a short voice input in Notepad or a browser text field.

v0.1.3-preview

This release is a full stability, privacy, and distribution upgrade covering local GPU inference, recording, model lifecycle, safe text entry, recovery, and onedir packaging, with local validation completed.

New Features:

📋 Single instance lock — Prevents running two copies at once
🔒 Local loopback transport — Uses a random port and a fresh token per run

Improvements:

💬 Friendly error messages — Errors now show clear Chinese explanations (e.g., "microphone not found" instead of raw error codes)
🖥️ GPU error dialog — When NVIDIA GPU is not found, a clear popup explains the problem instead of silently crashing
⚙️ Settings panel — Increased height to fit all new settings without scrolling issues
📝 Default config — Added missing P2 feature fields (auto_start, punctuation, VAD, spacing)

Bug Fixes:

Fixed tray "Settings" menu not working (was looking for a nonexistent window)
Fixed process_manager.rs comment missing "memory" strategy
Fixed build_release_zip.ps1 hardcoded version string

v0.1.1-preview