Music_Genre_Classification

Music_Genre_Classification

Music genre classification is one of the most basic yet fascinating DL/ML project. People use different approaches ranging from basi classification techniques like Survey and questionnaire to highly advanced combinations of Machine Learning Classifiers. The script that follows is a very basic implementation of one such model, revolving around Deep Learning, Sequential Model.

This script will not only enable you to setup a basic classification system, but will also help in feature extraction, dataset division, training and testing. Please be advised, that the aim is not to setup a really advanced classification system, but rather a simple setup, using a renowned dataset, to predict the genre of a musical piece using minimal features.

Dependency

The dependencies are summed uup in requirements.txt.

A separate list is also available below:

• librosa
• csv
• Keras
• TensorFlow
• numba
• numpy
• pandas
• scikit-learn
• scipy

Considering there are some big libraries here, it is recommended to setup a virtual environment first. This link might help.

Dataset

We have used GTZAN dataset for our task. It is publicly available, here is the link. For more information about the Dataset, click here.

Feature Extraction

LibRosa along with SciPy has been used for Feature Extraction.

As this is a much simpler approach, only 6 Features have been used along with their deltas, taking the total number of useful features to 26.

Here is a list of the Features.

• chroma stft
• rmse
• spectral centroid
• spectral bandwidth
• rolloff
• zero crossing rate

For those interested, this and this link will provide explanations about these and many more useful Audio Features. The features extracted are contained in a CSV, which can be used by anyone if the user wants to test accuracy of another classifier using the same features extracted.

Setup

1. A virtual environment (recommended)
2. pip install -r requirements.txt
3. Extract Features using feature_extraction.py, it is recommended to use the CSV file, rather than downloading the whole Dataset if same features re being extracted.
   • If different features are to be extracted, or the user wants to extract features himself only, download the dataset (use links provided above). Feed the path, proper headers and extract features.
4. The features extracted are saved in a csv file, the same file is used for training, the genre_classification.py script contains proper comments to help the user.

Sample Output

7/7 [==============================] - 2s 5ms/step - loss: 2.2042 - accuracy: 0.1935
Epoch 2/20
7/7 [==============================] - 0s 5ms/step - loss: 1.8557 - accuracy: 0.3668
Epoch 3/20
7/7 [==============================] - 0s 5ms/step - loss: 1.6626 - accuracy: 0.3710
Epoch 4/20
7/7 [==============================] - 0s 5ms/step - loss: 1.4863 - accuracy: 0.4496
Epoch 5/20
7/7 [==============================] - 0s 5ms/step - loss: 1.3252 - accuracy: 0.5336
Epoch 6/20
7/7 [==============================] - 0s 5ms/step - loss: 1.2093 - accuracy: 0.5683
Epoch 7/20
'
'
'
'
'
Epoch 18/20
7/7 [==============================] - 0s 5ms/step - loss: 0.5832 - accuracy: 0.8097
Epoch 19/20
7/7 [==============================] - 0s 6ms/step - loss: 0.5832 - accuracy: 0.8078
Epoch 20/20
7/7 [==============================] - 0s 6ms/step - loss: 0.5296 - accuracy: 0.8529
7/7 [==============================] - 0s 3ms/step - loss: 1.0282 - accuracy: 0.6750

Here is a brief summary of the model:

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 256)               6912      
_________________________________________________________________
dense_1 (Dense)              (None, 128)               32896     
_________________________________________________________________
dense_2 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_3 (Dense)              (None, 10)                650       
=================================================================
Total params: 48,714
Trainable params: 48,714
Non-trainable params: 0
_________________________________________________________________

In order to access the pre-trained model, click here

Note

The average accuracy obtained is 67.50%, which is obviously not a really good number, but one needs to understand that with only 6 Major features and a simple implementation of Sequential Model, without any feature selection or rejection, one can't hope for a better accuracy. One must understand that GTZAN is a really old dataset and after going through multiple research papers, the best number I was able to find was ~85%.

This code provides a really basic implementation of Deep Learning and Music Genre classification, users are free to manipulate the code and find better combination themselves.

P.S. - Even this implementation can attain higher accuracies with minor tweaks.

275/275 [==============================] - 5s 20ms/step - loss: 0.7065 - acc: 0.7793 - val_loss: 0.9531 - val_acc: 0.7172
Epoch 499/500
274/275 [============================>.] - ETA: 0s - loss: 0.7031 - acc: 0.7777Epoch 1/500
275/275 [==============================] - 5s 19ms/step - loss: 0.7019 - acc: 0.7780 - val_loss: 0.9383 - val_acc: 0.7258
Epoch 500/500
274/275 [============================>.] - ETA: 0s - loss: 0.6644 - acc: 0.7929Epoch 1/500
275/275 [==============================] - 5s 20ms/step - loss: 0.6635 - acc: 0.7932 - val_loss: 0.9442 - val_acc: 0.7273

Here the attained accuracy is closer to 80%.

Author(s)

Made by Vybhav Chaturvedi

If you like this script, I have made Repos related to Music Classification and Music Human Emotion mapping , Music Feature Classification.