Audio Segmentation Using K-means Clustering

Table of contents

• General info
• Library
• Module

General Info

My project to separate audio with the noise using k-means clustering based on the idea from the paper L. Marchegiani and I. Posner, "Leveraging the urban soundscape: Auditory perception for smart vehicles," 2017 IEEE International Conference on Robotics and Automation (ICRA), 2017, pp. 6547-6554, doi: 10.1109/ICRA.2017.7989774 [1].

Package Requirement

• numpy
• cv2
• IPython
• matplotlib
• skimage
• sklearn
• pandas
• argparse
• librosa
• scipy

Description

The system consists of two main scripts:

• spec.py
  • This script create a grayscale mel-spectrogram images with bandpassfilter from an audio using the datalist of a csv that you already created.
• k-means.py
  • This script cluster the grayscale mel-spectrogram images into several cluster then make a binary mask based on the threshold that you have decided.

Run

• Run spec.py first to create grayscale mel-spectrogram images, then run k-means.py to create the binary mask.
• Individual implementation using Jupyter Notebook is also provided on note_masking.ipynb and note_filter.ipynb.

Check results

You can check the spectrogram and k-means result in the image files in the directory Ground_truth/mel_spec and Ground_truth/mask respectively.