Audio Feature Selection and Classification

For this project, we wrote a genetic algorithm to find the near-optimal feature vectors for our audio classifiers. We have two classifiers, which each use a neural net and a KNN model to classify audio files as 1) music or non-music and 2) into separate musical genres. We selected a set of common audio features used to describe qualities of the music such as brightness, tempo, percusiveness, noise, etc. We then run a genetic algorithm on this initial population and evaluate each of the individuals (feature vectors) by training the models with their data and getting their accuracy. After n number of generations, we return the fittest individual, i.e. the feature vector with the greatest accuracy.

Index of the repositiory:

1. /classifiers -- contains MOClassifier, a wrapper for KNN algorithm
2. /experimental-code -- you can ignore this directory
3. extract_features.py -- given a list of feature names, returns feature vector using Librosa calls to extract features
4. MusicAgent.py -- run this to segment inputted file, run genetic algorithm, and output results for
5. select_features.py -- contains genetic algorithms for KNN and neural nets (optimizing parameters feature selection)
6. /data -- contains .csv files containing features extracted for each sample, for training the models

See code_walkthrough.mp4 for a demonstration of code