sandcobainer / drum-sample-clustering

Drum Samples Clustering, Audio feature extraction and clustering audio files using data visualization and dimensionality reduction (PCA).

Interpolating-Drum MIR Semester project, Fall'19

document created by: Sandeep Dasari
last updated: 2019, October 28
last updated by: Sandeep

Contents

• Demo
• Overview
• Dependencies Overview
• Folder Contents

Video Demo

Overview

Interpolating-Drum

Inspired from Google Experiments' The Infinte Drum Machine, this project seeks to take the idea of 2 Dimensional visualization of higher dimensional audio data. Unlike t-SNE that maps high dimensional STFT data to 2-D/3-D , this project focuses on using simpler musical descriptors or popular music information retreival features like

• Spectral centroid
• RMS
• Spectral Flux
• Attack
• Decay etc.

The reason for simplifying the dimensions is to allow the user to interactively sort a large pool of audio samples forming a general sense for these features as filters of sonic quality. The dataset was parsed and processed to extract features in Python. Later a Principal Component Analysis is run to reduce dimensions and perform a K-Means on the sample library. The results from PCA are sent to Max via OSC for interactive audio plotting of the one-hit samples.

Dependencies Overview

• Max 8 for visualization and sonofication
• python-osc
• matplotlib
• sklearn
• essentia
• py-dub
• librosa
• pandas
• numpy

Folder Contents

• /samples contains all the 160 one hit acoustic drum kit sample
• /main.py is the python file that parses samples and sends PCA co-ordinates to Max/MSP
• /viz.maxpat contains the Max patch that receives OSC, plots audio samples and plays them in real time.
• /pca.mov is an initial demo video of the application