wa-transcriber 1.3.5

Automated local WhatsApp voice note transcriber

🎙️ WhatsApp Auto-Transcriber

A lightning-fast, privacy-focused tool that automatically transcribes incoming WhatsApp voice notes on macOS and Windows.

Why? WhatsApp's native transcription is often unavailable or slow. This tool provides a superior, 100% local alternative that works instantly the moment a voice note is downloaded, leveraging your hardware (GPU/NPU) for maximum speed.

[!IMPORTANT] Desktop Only: This tool requires the official WhatsApp Desktop App (Store Version). It monitors the internal file system for new .opus files.

✨ Features

• ⚡️ Hardware Accelerated:

  • macOS: Native Metal Performance Shaders (MPS) for M1/M2/M3/M4 chips.
  • Windows: Native NVIDIA CUDA support for GeForce RTX cards.

• 📂 Intelligent Monitoring: Uses a threaded Watchdog observer to detect voice notes instantly without locking your system.

• 🏥 System Health Check: Includes a utility to analyze your RAM/VRAM and suggest the optimal AI model size to prevent crashes.

• Smart Maintenance: Automatically cleans up AI models that haven't been used in 3 days (configurable) to save disk space.

• Startup Backfill: Scans for missed voice notes from the last hour upon launch to ensure nothing is lost.

• 📋 Auto-Clipboard & Logging:

  • Text is copied to clipboard (Cmd+V / Ctrl+V).
  • Transcripts are saved to daily logs (e.g., 2026-01-27_daily.log).

🚀 Prerequisites

1. Install FFmpeg

Whisper relies on FFmpeg for audio processing.

• macOS: brew install ffmpeg
• Windows: choco install ffmpeg (or download binaries from ffmpeg.org and add to PATH).

2. Python 3.10+

Ensure you have a modern Python version installed. You can download it from python.org.

🛠️ Installation

From PyPI (recommended)

Install the latest release with pip (use a venv if you like):

pip install wa-transcriber

Or install as an isolated CLI with uv:

uv tool install wa-transcriber

The package is published on PyPI. Dependencies include PyTorch and Whisper, so the first install may take a while and use significant disk space.

From source

Clone the repository and install from the checkout (for development or unreleased changes). Use a venv if you prefer.

git clone https://github.com/jpxoi/wa-transcriber.git
cd wa-transcriber
pip install .

For an editable install while hacking on the code: pip install -e .

If you use uv in the repo:

git clone https://github.com/jpxoi/wa-transcriber.git
cd wa-transcriber
uv sync

Run CLI commands with uv run wa-transcriber ... (for example uv run wa-transcriber health).

For day-to-day use of a stable build, prefer From PyPI above.

Run the Health Check (Crucial)

Before running the main program, run the included health check tool. This script analyzes your System RAM (CPU/MPS) or VRAM (NVIDIA) and calculates exactly which Whisper model your computer can handle safely.

wa-transcriber health

• ✅ Verifies FFmpeg installation.
• ✅ Detects Hardware Acceleration (CUDA vs MPS vs CPU).
• ✅ Calculates memory overhead and suggests the best MODEL_SIZE.

⚙️ Configuration

Run the interactive setup wizard to configure the application:

wa-transcriber setup

This will guide you through:

• Model Selection: choosing the appropriate Whisper model size.
• Language: setting a preferred language or using auto-detection.
• WhatsApp Path: automatically detecting or manually specifying the WhatsApp Media folder.
• Hardware Limits: configuring RAM and VRAM usage limits.

To view your current configuration at any time, run:

wa-transcriber config

Configuration is stored in config.json in ~/.wa-transcriber/.

Model Selection Reference

The setup wizard will suggest a model based on your hardware, but you can choose any of the following:

Model	VRAM/RAM Req	Speed	Accuracy	Best For
tiny	~1 GB	⚡️ Instant	Low	Older laptops
base	~1 GB	🚀 Very Fast	Decent	Quick snippets
small	~2 GB	🏃 Fast	Good	General usage
medium	~5 GB	⚖️ Balanced	Great	Professional use
turbo	~6 GB	🏎️ Optimized	Excellent	M1/M2/M3 & RTX 3060+
large-v3	~10 GB	🐌 Slow	Perfect	Heavy accents / Noisy audio

Advanced Settings

The following settings can be fine-tuned via the setup wizard or by manually editing config.json:

• Scan Lookback: Number of hours to check for missed files on startup.
• Model Cleanup: Automatically delete models unused for a set number of days.
• Memory Limits: Adjust how aggressively the script uses System RAM or GPU VRAM.

🏃 Usage

Run the main script. It will initialize the model and start watching the folder.

wa-transcriber

Workflow:

1. Script loads (shows a progress bar for model loading).
2. Startup Scan: Checks for missed files from the last hour.
3. "👀 Watching Folder" message appears.
4. Receive a voice note in WhatsApp Desktop.
5. Instant Result:

• Console shows: ⚡️ [WORKING] Processing: audio_file.opus
• Then: ✅ [DONE] Transcript: Hello world...
• Clipboard: Updated automatically.

🛠️ CLI Reference

Command	Description
wa-transcriber	Starts the main transcription service.
wa-transcriber setup	Runs the interactive configuration wizard.
wa-transcriber health	Runs system diagnostics and hardware checks.
wa-transcriber config	Displays the current configuration.
wa-transcriber reset	Resets the application by removing all user data and configuration.
wa-transcriber logs audio	Shows the last 50 lines of transcribed audio logs.
wa-transcriber logs app	Shows the last 50 lines of application logs.

❓ Troubleshooting

• "Clipboard unavailable": On Linux, you may need xclip or xsel. On Windows/Mac, this usually works out of the box.
• "CUDA out of memory": Run wa-transcriber health and switch to a smaller model (e.g., from large to medium).
• Script doesn't trigger: Ensure "Media Auto-Download" is ON in WhatsApp settings, or manually click the download arrow on the voice note.

🤝 Contributing

Pull requests are welcome. For major changes, please open an issue first.

📄 License

© 2026 Jean Paul Fernandez. Licensed under GPLv3.