Deepfake Audio Detector

A fully local, Celery-backed WhatsApp bot that detects AI-generated voice notes and audio clips using PyTorch, AASIST, and Wav2Vec2 models.

This guide explains how to spin up the entire infrastructure locally on your own computer.

Prerequisites

1. Python 3.10+
2. FFmpeg (sudo dnf install ffmpeg or sudo apt install ffmpeg)
3. Ngrok Account (to expose your local bot to Meta)
4. Upstash Account (for a free, serverless Redis queue)
5. Meta Developer Account (to use the WhatsApp Business API)

---

1. Installation

Clone the repository, create a virtual environment, and install the dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Download the AASIST model weights (if you haven't already):

git clone https://github.com/clovaai/aasist /tmp/aasist
cp /tmp/aasist/models/weights/AASIST.pth ./weights/AASIST.pth
cp /tmp/aasist/models/AASIST.py ./models/AASIST.py

---

2. Configuration & Secrets

Copy the example environment file to .env:

cp .env.example .env

Open .env and fill in the following three critical values:

1. WA_TOKEN: Go to your Meta Developer Dashboard -> WhatsApp -> API Setup. Click Generate access token. (Note: This token expires every 24 hours. If your bot stops replying with HTTP 401 Unauthorized, you need to generate a new one and paste it here).
2. PHONE_NUMBER_ID: Found on the same Meta API Setup page.
3. REDIS_URL: Go to Upstash.com, create a free Redis database, copy the URL, and ensure it begins with rediss://.

---

3. Running the Infrastructure (The 3 Terminals)

Because this bot is built for heavy AI processing, it is split into a web server (FastAPI) and a background worker (Celery). You need three separate terminal windows open simultaneously.

Terminal 1: The FastAPI Web Server

This server catches incoming messages from Meta and instantly puts them into the Upstash Redis queue.

source .venv/bin/activate
uvicorn webhook.main:app --host 0.0.0.0 --port 8000 --env-file .env

Terminal 2: The Celery Worker

This worker pulls messages out of Redis, downloads the audio, runs the PyTorch models, and sends the final verdict back to WhatsApp.

source .venv/bin/activate
celery -A processor.tasks worker --loglevel=info

Terminal 3: The Ngrok Tunnel

This exposes your local Port 8000 to the public internet securely so Meta can reach it.

ngrok http 8000

---

4. Connecting Meta to your Bot

Once your 3 terminals are running, look at Terminal 3 (Ngrok) and copy the Forwarding URL (it looks like https://random-words.ngrok-free.dev).

1. Bypass Ngrok Warning: Open a web browser and paste your Ngrok URL. Click the blue "Visit Site" button. (If you skip this, Meta will fail to verify the webhook).
2. Go to the Meta Developer Dashboard -> WhatsApp -> Configuration.
3. Under the Webhook section, click Edit.
4. Callback URL: Paste your Ngrok URL and append /webhook to the end. (e.g., https://random-words.ngrok-free.dev/webhook).
5. Verify Token: Type in the exact VERIFY_TOKEN you set in your .env file.
6. Click Verify and Save.
7. Directly below that, in the "Webhook fields" section, click Manage and subscribe to messages.

---

5. Usage

You are completely live! Message your Meta Test Phone Number via WhatsApp.

• You can record a Voice Note.
• You can attach a .wav or .ogg file as a Document.

The bot will automatically download, convert, analyze, and reply with a trust score.