Electronic music styles Neural Net

The repository contains the model, a convolutional neural network, which was trained to distinguish 12 different styles of modern electronic music, trained to classify 12 classes. Keras model .h5 and super handy Jupyter Notebook .ipynb as a guide with details are provided. Starring Tensorflow as backend.

Data

The network was trained on a private dataset of labeled raw audio, processed as Mel Spectrograms. Data will not be provided. 12 electronic music styles (i.e. classes) are:

• Ambient / Electronic (downtempo, abstract);
• Brostep (Dubstep, as it more popularly known, songs of Skrillex, MUST DIE!, Kill The Noise and the others);
• Deep House (proggy and techy, from the likes of Anjunadeep);
• Drum & Bass (techstep, Neurofunk, liquid funk among the others);
• EDM & Big Room (SHM, Hardwell, Martin Garrix and the others);
• Electro & Future House;
• Trance (Uplifting, psy, stadium);
• Halftime (Drum & Bass subgenre with the same tempo, but with a half less drum kicks);
• House / Nu Disco / Disco;
• Progressive House (also, its modern form, mostly, Yotto, Jeremy Olander, Fehrplay, Komytea);
• Techno / Tech House;
• Trap

A potential problem with the model reasoning is objectivity since the data was labeled manually mostly. The goal of choosing exactly them is to check ability to distinguish relatively similar styles. Like, distinguish the style of Trap from Halftime.

How good is it?

You can find the full description of model inference in handy_guide.ipynb.

We evaluate our model in terms of Precision @ K:

• Precision @ 1 = 67% (i.e. accuracy, top one class)
• Precision @ 2 = 83% (top two classes)
• Precision @ 3 = 89% (three top classes)

Requirements

FFMPEG | ffmpeg version N-91266-g8c20ea8ee0 | https://www.ffmpeg.org/

Python 3.6.5 package requirements:

tensorflow-gpu==1.8.0
librosa==0.6.1
keras==2.1.6
mutagen==1.40.0
kapre
numpy==1.14.3
pandas==0.23.0
matplotlib==2.2.2
notebook==5.5.0

All together in requirements.txt. Sitting in the directory, execute pip3 install -r requirements.txt

All the sources of priceless knowledges

Transfer learning for music classification and regression tasks | https://arxiv.org/abs/1703.09179 AUTOMATIC TAGGING USING DEEP CONVOLUTIONAL NEURAL NETWORKS – Keunwoo Choi, et al ｜ https://arxiv.org/abs/1606.00298 Recommending music on Spotify with deep learning | http://benanne.github.io/2014/08/05/spotify-cnns.html Japan company QOSMO AI DJ Project | http://qosmo.jp/projects/2016/11/21/ai-dj-human-dj-b2b/