Voclaude

Voice input anywhere — local-first, GPU-accelerated speech-to-text that runs in your system tray.

Press a hotkey, speak, and your words are instantly transcribed and copied to your clipboard. All processing happens locally using Qwen3-ASR via pure Rust inference (no Python needed).

Features

• Single binary — pure Rust, no Python or external runtimes
• Local-first — all transcription happens on-device, nothing leaves your machine
• GPU accelerated — CUDA inference at ~30x realtime on an RTX 3060 (Q8 weights, fused kernels)
• Transcribes while you speak — the recording is cut into segments at pauses and transcribed in the background; after you stop, only the last few seconds remain, so the text is on the clipboard within about a second
• Never loses a word — audio is streamed to disk and synced every 5 s; every finished segment's text is journaled; a crash or power cut resumes from the last finished segment on relaunch, and the tray has "Recover Last Recording"
• Tiny when idle — inference runs in a child process that exits after 60 s idle: 0 VRAM and ~20 MB of RAM for the tray
• Lazy loading — model downloads on first use; loads while you are still speaking
• System tray — lives in your tray, stays out of your way
• HUD overlay — recording/transcribing status appears as a small overlay
• History window — browse, search, and copy past transcriptions (Ctrl+Shift+H)
• Global hotkey — configurable hotkey (default: F4) works from any application
• Session recovery — if the app crashes mid-transcription, it recovers on restart

Quick Start (download)

1. Download voclaude-vX.Y.Z-gpu-setup.exe from the Releases page and run it. It installs for the current user (no admin prompt), can start with Windows, and launches the app. A tray icon appears. Prefer no installer? voclaude-vX.Y.Z-<hash>-gpu.zip unzips anywhere; run voclaude.exe from the folder.
2. Press F4, speak, press F4 again. The text is on your clipboard; paste it with Ctrl+V.

On the first run the app downloads the model (about 3.4 GB) from Hugging Face; the tray/HUD shows "Downloading ... first run only" while it does. Later starts load in about 1.5 s.

To start it with Windows: press Win+R, type shell:startup, and put a shortcut to voclaude.exe there.

Requirements (download)

• Windows 10/11, 64-bit.
• Best experience: an NVIDIA GPU from the GTX 16xx / RTX 20xx generation or newer with 6 GB or more VRAM, and a driver from 2025 or newer (CUDA 13 runtime; nothing else to install — the zip bundles the CUDA DLLs). On a GPU with less memory the app automatically switches to the smaller 0.6B model.
• No NVIDIA GPU: it still works on the CPU with Q8 weights: about 2x realtime for the 1.7B model and 3.4x for 0.6B on a desktop CPU (see docs/bench/README.md), which keeps up with speech because segments are transcribed while you talk. Needs a CPU with AVX2 (2013+).
• Microphone. Windows: Settings → Privacy → Microphone → allow desktop apps.

Quick Start (build from source)

Prerequisites

• Rust 1.70+ (rustup.rs)
• NVIDIA GPU with CUDA Toolkit 12.x+ (for GPU mode)
• ~4.5 GB disk space for the model (downloaded automatically on first run)

Build and Run

git clone https://github.com/matvei77/Voclaude.git
cd Voclaude
cargo build --release

# Windows
target\release\voclaude.exe

# Linux
./target/release/voclaude

CPU-only Build

If you don't have an NVIDIA GPU:

cargo build --release --no-default-features --features cpu

Packaging for Distribution

To build a distributable zip with CUDA DLLs bundled:

.\package.ps1          # GPU build -> dist\voclaude-vX.Y.Z-gpu.zip
.\package.ps1 -Cpu     # CPU build -> dist\voclaude-vX.Y.Z-cpu.zip

The GPU zip includes voclaude.exe, the required CUDA DLLs, and config.example.toml. Recipients only need NVIDIA GPU drivers — no CUDA Toolkit or Rust toolchain.

Linux Dependencies

# Ubuntu/Debian
sudo apt install libasound2-dev libgtk-3-dev libayatana-appindicator3-dev \
    libxkbcommon-dev libxdo-dev pkg-config cmake clang

Usage

1. Launch — Voclaude appears in your system tray
2. Press F4 to start recording — a HUD overlay appears top-right
3. Speak — your voice is captured locally
4. Press F4 again to stop — segments were already transcribed while you spoke; the last one finishes and the HUD shows "Copied to clipboard"
5. Paste (Ctrl+V) — transcribed text is in your clipboard

Keyboard Shortcuts

Tray Menu

Right-click the tray icon for: Show History, Open Last Transcript, Open Transcripts Folder, Recover Last Recording, Settings, Quit.

Local API (for scripts and agents)

While the app runs it serves http://127.0.0.1:7770/v1/audio/transcriptions, an OpenAI-compatible transcription endpoint that decodes audio and video files (wav, mp3, flac, ogg, m4a/mp4/mov, mkv; anything else via ffmpeg if installed) and uses the same loaded model as dictation:

voclaude transcribe meeting.mp4            # prints the text
curl -F file=@clip.wav http://127.0.0.1:7770/v1/audio/transcriptions

See docs/API.md. Disable with api_enabled = false.

Configuration

Config is created automatically on first run.

Location: %APPDATA%\voclaude\Voclaude\config\config.toml

hotkey = "F4"
history_hotkey = "Ctrl+Shift+H"
add_trailing_space = true
capitalize_first = true
idle_unload_seconds = 60
use_gpu = true
model = "Qwen/Qwen3-ASR-1.7B"
quantization = "q8_0"        # "none" for F16 weights
streaming = true             # transcribe segments while recording
segment_min_seconds = 20
segment_max_seconds = 45
segment_pause_seconds = 0.5

See config.example.toml for all options.

Supported Hotkeys

• Modifiers: Ctrl, Alt, Shift, Super/Win
• Keys: A-Z, 0-9, F1-F12, Space, Enter, Tab, Escape, Numpad keys, punctuation
• Examples: F4, Super+C, Ctrl+Shift+Space, Alt+V

Architecture

System Tray ─── Global Hotkey (F4)
    │                │
    │    ┌───────────▼───────────┐
    │    │    Audio Capture      │
    │    │  cpal + 16kHz resamp  │
    │    └───────────┬───────────┘
    │                │
    │    ┌───────────▼───────────┐
    │    │   Qwen3-ASR Candle    │
    │    │   Pure Rust inference  │
    │    │   CPU or CUDA          │
    │    │   ~4.5 GB model        │
    │    └───────────┬───────────┘
    │                │
    │    ┌───────────▼───────────┐
    │    │     Clipboard         │
    │    │  Formatted + copied   │
    │    └───────────────────────┘
    │
    ├── HUD Overlay (recording/transcribing status)
    └── History Window (searchable transcript archive)

Troubleshooting

Model download is slow

The Qwen3-ASR model (~4.5 GB) is downloaded from Hugging Face on first use. Ensure you have a stable internet connection and sufficient disk space.

No audio input detected

Check that your microphone is set as the default input device in your system sound settings.

CUDA errors

• If using the prebuilt zip: make sure you have an NVIDIA GPU and up-to-date drivers (nvidia-smi should work)
• If building from source: ensure CUDA Toolkit 12.x+ is installed
• Try CPU mode: rebuild with --no-default-features --features cpu

Hotkey not working

• Make sure no other application has registered the same global hotkey (F4)
• Try a different hotkey in the config file (e.g. hotkey = "Ctrl+Shift+Space")

Linux: Hotkey not working on Wayland

Global hotkeys may require X11 compatibility:

GDK_BACKEND=x11 ./target/release/voclaude

License

MIT