Zecho

Voice to clipboard, instantly. Speak naturally, get clean text -- no subscriptions, no cloud, no ads. Everything runs locally on your machine.

Hold Shift+Fn to record, release to get polished text copied to your clipboard. That's it.

How It Works

1. Speak -- Hold Shift+Fn and talk naturally. Corrections, pauses, and filler words are all fine.
2. AI cleans -- Local speech recognition (Bark or Whisper) transcribes your speech. A small local model removes filler words, fixes grammar, and resolves self-corrections ("I want red, uh, no actually blue" becomes "I want blue").
3. Paste -- Clean text lands in your clipboard. Paste anywhere.

Features

• Fully local -- No data leaves your machine. No accounts, no network calls, no telemetry.
• Shift+Fn recording -- Hold to record, release to transcribe. Double-tap to lock recording. Switch to plain Fn in Settings.
• AI text cleanup -- Removes "um," "uh," filler words, and resolves self-corrections using a small local model.
• Writing styles -- Formal, Casual, or Very Casual. Custom prompts for fine-tuning.
• Cleanup levels -- None, Light, Medium, or High -- control how aggressively AI edits your text.
• Clipboard history -- Every transcription is saved with before/after comparison (raw transcription vs. cleaned text).
• macOS native floating panel -- Hover and interact without stealing focus from your active window.
• Lightweight -- Built with Rust and Tauri v2. ~3 MB app overhead.
• Auto updates -- New versions are detected and applied automatically.

Getting Started

Prerequisites

• macOS 12+
• Rust toolchain (rustup)
• Tauri CLI v2 (cargo install tauri-cli --version "^2")

Setup

# Clone the repo
git clone https://github.com/dzearing/zecho.git
cd zecho

# Download the default models (Whisper + Spokenly Cleanup)
./scripts/download-models.sh

# Run in development mode
cargo tauri dev

The model download script places files in ~/Library/Application Support/zecho/models/. The app will also download models on demand through its settings UI.

Project Structure

ui/              # Frontend (HTML/CSS/JS)
src-tauri/       # Rust backend (Tauri v2)
  src/
    audio.rs       # Audio capture (cpal)
    transcribe.rs  # Speech-to-text engine selection (Whisper / Bark)
    bark.rs        # Bark 3.0 on Core ML (Neural Engine)
    cleanup.rs     # LLM text cleanup
    hotkey.rs      # FN key listener (macOS)
    history.rs     # Clipboard history storage
    settings.rs    # User preferences
    models.rs      # Model management & downloads
    macos_panel.rs # Native floating panel
scripts/         # Setup scripts
docs/            # Landing page (GitHub Pages)

Download

Get the latest release: github.com/dzearing/zecho/releases/latest

Speech models

Zecho uses Whisper by default. On macOS 15+ with Apple silicon, English users can opt into Bark 3.0 in Settings. It runs on the Neural Engine and is faster and more accurate than Whisper base.en, but it's English-only, and after download it takes a few minutes to prepare for your Mac.

Bark 3.0 is Underdog Bark 3.0 by Conway Research, derived from nvidia/parakeet-tdt-0.6b-v3 (© NVIDIA Corporation) and moondream/parakeet-ultra (© Moondream). It's used under CC-BY-4.0; Zecho downloads the model unmodified, along with its NOTICE and LICENSE.

Text cleanup defaults to Spokenly Cleanup 1 (Apache-2.0), a small model trained only for dictation cleanup. It follows the writing style but not the cleanup level or custom instructions; pick a Gemma model in Settings for those.

License

Apache 2.0