Music-Genre-Classification (GTZAN Dataset)

Prerequisite

• Python 3.6.8
• CUDA 9.0 (Follow these: https://blog.quantinsti.com/install-tensorflow-gpu/ )
• Sublime Text (Optional)

Package Required

• numpy (For Numerical Computation)
• librosa (For dealing with Audios)
• tqdm (For showing loading/progrss bar)
• sklearn (For Splitting the data into train, valid and test, and Confusion Matrix)
• keras (For Deep learning: CNN and VGG16)
• matplotlib (For showing the graph of train and valid with loss and accuracy)
• collections (For storing the genres corresponding names to showing in confusion matrix)
• itertools (For iterating the elements)
• pickle (For storing the data into hard drive, so that we don't need to compute again and again. Just call the pickle file and load)

System Requirement

• 16GB RAM
• Nvidia 4GB RAM

Dataset

GTZAN Genre Collection (Download Link: http://marsyas.info/downloads/datasets.html )

Steps of Execution:

1. Read_File.py : This file perform reading the audio files with labels corresponding to number of genres.

2. Audio_Segment.py : This file perform clipping of audio into small segment/clip of duration depending on window size and overlap.

3. Feature_Extraction.py : This file contain feature extraction techniques (STFT, Melspectrogram and MFCC). Whichever you want to extract, just change the name of the function 'to_stft' instead.

4. Split_Data.py : This file perform splitting of data into train, valid and test data.

5. CNN_Model.py : This file contain CNN model. (Also include CNN + RNN in comment section) So that you can train with RNN if required.

6. CNN_BiDirectional.py : This file is for CNN+BiRNN Model.

7. VGG16_Model.py : This file contain VGG16 model. (Also include VGG16 + RNN in comment section) So that you can also train with RNN also if required.

8. VGG16_BiDirectional.py : This file is for VGG16+BiRNN Model.

Other files are just for EDA:

Waveform.py : To show the plot of wave for the audio

Plot_Audio.py : To plot the spectogram of segments of the audio. (You can uncomment the code if you want to see STFT, Melspectrogram or MFCC)

Plot_CM.py : Create a module to print the Confusion Matrix

Result

See the result with every model with different feature extraction in 'Result and Output' Folder

• train, valid and test loss and accuracy
• test confusion matrix

Description

You can study the description/thesis about this project in 'Major Thesis_(Final).pdf'

Publication

Faiyaz Ahmad, Sahil,'Music Genre Classification using Spectral Analysis Techniques With Hybrid Convolution-Recurrent Neural Network',International Journal of Innovative Technology and Exploring Engineering (IJITEE), Volume-9, Issue-1, Novemeber, 2019

Link: https://www.ijitee.org/wp-content/uploads/papers/v9i1/A3956119119.pdf