The aim of this project is the discover what combination of audio features gives the best performance with electronic versus organic source classification. Source recognition is treated as a binary classification problem, with a sound represented as either orginating from a live in-person source or an electronic source. Features extracted by Essentia and LibROSA, tools for audio analysis and audio-based music information retrieval, were used.
- audio-ml-extraction.ipynb contains ML classifiers implemented on existing dataset
- essentia_all_features.ipynb contains methods for extracting relevant audio features using Essentia
- librosa_feature_extraction.ipynb contains methods for extracting relevant audio features using LibROSA
- librosa_features.csv = example output of LibROSA extraction
- features_mfccaggr.sig = example output of signal-level Essentia extraction
- features_frames_aggr.sig = example output of frame-level Essentia extraction
- /audio = current audio dataset
- /tutorials = working Essentia examples