andriosrobert / deeptrolldetector

Deep troll (Gemidao do Whatsapp) detector

Tired of getting caught in difficult situations after playing some Whatsapp content with a troll in it?

Let's put deep learning to work for the good!

Deep troll uses a deep learning model that identifies whether an audio contains the Gemidao troll (AAAWN OOOWN NHAAA AWWWWN AAAAAH).

The model uses a RNN-GRU architecture, using convolutions on audio spectrograms to extract features.

Setup

1. Use the requirements.txt file to install the dependencies (preferably using virtualenv)

git clone https://github.com/andriosr/deeptrolldetector.git
cd deeptrolldetector
virtualenv deeptroll_env
source deeptroll_env/bin/activate
pip install -r requirements.txt

2. Download the training set and pre-trained models at:

https://drive.google.com/file/d/1wBfrJ7UPC0BLJZlJRlSZgSHxC_YButiF

3. Extract the downloaded file in the project folder, setting its name to data

4. Start Jupyter, open the Deep_Trool_Detector.ipynb and let the fun begin

jupyter notebook