A music classifier in Python for CPSC 473.
Written in Python 3.6
I cut all songs down into 30 second .wav files using ffmpeg. I then used Marsyas to collect Mel-frequency cepstrual coefficients, from the 30 second clips. MFCCs are coefficients that represent the short-term power spectrum of a sound. Putting the MFCC vectors into a SVM I was able to classify different genres of music.
I've included test data with a different number of genres so that you do not have to collect your own.
After running the classifier with different number of genres we can see that after 4-5 genres the classifier starts to decline in accuracy.
| Number of Genres | % Correct | % Wrong |
|---|---|---|
| 2 | 92.5% | 7.5% |
| 3 | 68% | 32% |
| 4 | 60% | 40% |
| 5 | 62% | 38% |
| 7 | 46% | 53% |
| 10 | 43% | 56% |
pip3 install -r requirements.txt
Run the code
python3 Main.py <Path/To/.arff>
To generate a .arff file I used Marsyas to get the Mel-frequency cepstral coeffcients in a .arff file, in a single vector.
problem with tkinter not being installed? try
sudo apt-get install python3-tk
Use ffmpeg to convert mp3 to 30 second wav
ffmpeg -ss 00:00:30 -t 00:00:30 -i <song.mp3> <song.wav>
[1] R. Griesmeyer, "Music Recommendation and Classification Utilizing Machine Learning and Clustering Methods." (2011). Florida State University.
[2] M. Haggblade, Y. Hong, K. Kao "Music Genre Classification" (2011).
[3] T. Li, L. Li "Music Data Mining: An Introduction" (2010)
[4] Y. Costa, L. Oliveira, C. Silla Jr. "An evaluation of convolutional Neural Networks for music classification using Spectrograms" (2016)