AI vs Human song classifier. The idea is to use cpu rather than gpu.
https://huggingface.co/spaces/dkappe/AISong
pip install torch librosa matplotlib
and run against a directory with mp3 and wav files:
python classify2.py directory_path
The larger model uses two convolutional layers, two GRUs and a FC layer. Also a focal loss function.
Running against 166 human songs yields
163/166 = 98.2% correct
Running against 1075 AI songs yields
1054/1075 = 98% correct
Running against 1163 spotify AI songs yields
878/1163 = 75.5% correct
The latest net on spotify AI songs yields
1020/1163 = 87.7% correct
If you want to detect an artist that uses AI, put several albums worth of mp3's in a directory and run the above classifiers.