Music Genre Recogniton

This project is about classifying the genre of a song by using machine learning approach. The neural network used in this code is made up of CNN and LSTM. When compared this model with CNN-GRU, the CNN-LSTM approach performed better than the traditional CNN-GRU approach. The testing has been done on GTZAN dataset.

Dataset

For this project I have used the GTZAN dataset. This dataset has 1000 audio track and each is 30 sec long. This dataset consists of 10 genres. Download GTZAN here. It has the following genres:

• blues
• classical
• country
• disco
• hiphop
• jazz
• metal
• pop
• reggae
• rock

Built With

• Python3
• Keras (running tensorflow in the backend)

How it works

First I take each song from each genre one by one. To make a training set from audio files I convert audio files to their mel-spectograms. Mel-spectogram of an audio file may look like this:

I divided my dataset into three parts:

dataset = training set + test set + valid set

After converting to mel-spectogram this result is fed into the neural network structure of CNN-LSTM. The structure output is like below:

Model: "sequential_115"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_223 (Conv2D)          (None, 60, 169, 20)       520       
_________________________________________________________________
max_pooling2d_109 (MaxPoolin (None, 30, 84, 20)        0         
_________________________________________________________________
conv2d_224 (Conv2D)          (None, 26, 80, 50)        25050     
_________________________________________________________________
max_pooling2d_110 (MaxPoolin (None, 13, 40, 50)        0         
_________________________________________________________________
flatten_103 (Flatten)        (None, 26000)             0         
_________________________________________________________________
dense_127 (Dense)            (None, 20)                520020    
_________________________________________________________________
lambda_50 (Lambda)           (None, 20, 1)             0         
_________________________________________________________________
lstm_101 (LSTM)              (None, 512)               1052672   
_________________________________________________________________
dense_128 (Dense)            (None, 10)                5130      
=================================================================
Total params: 1,603,392
Trainable params: 1,603,392
Non-trainable params: 0

Libraries Used

• librosa -> details here.
• csv
• pandas
• numpy

Accuracy and Loss Plot

CNN-GRU accuracy = 50.30%, and

CNN-LSTM accuracy = ~61%

The CNN-LSTM VS CNN-GRU plot is like below:

References

• Recommending music on Spotify with deep learning https://benanne.github.io/2014/08/05/spotify-cnns.html
• K. Choi, G. Fazekas, K. Cho, and M. Sandler, “A tutorial on deep learning for music information retrieval,” arXiv preprint arXiv:1709.04396, 2017.
• Music Genre Recognition by Deep Sound http://deepsound.io/music_genre_recognition.html
• Using CNN and RNN for genre recognition by Medium https://towardsdatascience.com/using-cnns-and-rnns-for-music-genre-recognition-2435fb2ed6af
• K. Choi, G. Fazekas, M. Sandler, and K. Cho, “Convolutional recurrent neural networks for music classification,” in Proc. Int. Conf. Acoust, Speech, Signal Process., 2017
• Librosa on github - https://github.com/librosa/librosa