prateekbose / Music-Genre-Classification

Create a K-nearest neighbours (KNN) based model with Mel Frequency Cepstral Coefficients (MFCC) to accurately classify music samples into different genres based on various factors and also help reclassify broader classes to increase accuracy overall.

Music Genre Classification

The objective of the project is to implement a fairly accurate, easy to implement and computationally less taxing model to classify music into separate genres based on various factors.

Feature Extraction and Techninques used

There are two types of audio features: top-level and low-level. Top-level features give information such as genres, atmosphere, instruments, and so forth. Temporal domain features, frequency domain features, cepstral features, and modulation frequency features are examples of low-level features.

For the purpose of Feature Extraction, Mel-Frequency Cepstral Coefficients (MFCC) where extracted from the training set which were then used in the model.

Mel-Frequency Cepstral Coefficients (MFCC)

A signal's Mel-Frequency Cepstral Coefficients (MFCC) are a limited group of characteristics (typically 10-20) that simply represent the overall shape of a spectral envelope. To remove the noise, it performs a discrete cosine transform (DCT) on these frequencies. We only maintain a certain sequence of frequencies that have a high possibility of containing information when we use DCT.

K Nearest Nighbours

It is one of the clustering techniques that collects data points using similarity in proximity strategies. The mean distance between the locations is calculated using Euclidean distance. And the pattern continues to accumulate more points, eventually becoming a cluster. The K-nearest neighbours technique is used in the model as it has consistently shown the best results for this problem in many studies.

Dataset Used

For the project, the GTZAN dataset was utilised The GTZAN dataset is the most often used public dataset for music genre recognition assessment in machine learning research (MGR). In order to reflect a diversity of recording settings, the files were acquired in 2000-2001 from a number of sources, including personal CDs, radio, and microphone recordings.

The dataset can be downloaded from: https://www.kaggle.com/andradaolteanu/gtzan-dataset-music-genre-classification

Comparitive Results

Parameters/Methods	Existing Method 1	Existing Method 2	Propose Method
Dataset	3-Root and 9-Leaf Genre Dataset	GTZAN Dataset	GTZAN Dataset
Train-Test Ratio	90% train - 5% validation - 5% test	90% train - 10% test	90% train - 10% test
Techniques used	MIDI with NCD, KNN	MFCC, KNN	MFCC, KNN
Number of DistinguishableGenres	3	4	10
Test Accuracy	85%	80%	86%