The goal of this project is to generate audio playlists based on the precomputed audio features of a library of songs. For the analysis the Essentia library was used to extract features over the audio files contained in the MusAV dataset.
The MusAV Dataset contains 2,092 30 seconds track previews covering 1,404 genres, with arousal and valence annotations.
For the audio feature analysis the Essentia library (open-source library from Music Technology Group at UPF) was used. For each audio file the extracted features are: BPM, danceability, wether it is a vocal or instrumental prominent track, a value for arausal and valence and a genre prediction.
The script uses several Essentia algorithms to perform the audio analysis, including RhythmExtractor2013 for tempo analysis, Danceability for danceability analysis and several pretrained neural network models: voice_instrumental-musicnn-msd-1.pb for voice/instrumental classification, msd-musicnn-1.pb for arousal and valence analysis and discogs-effnet-bs64-1.pb for music style prediction.
Some other Essentia utilities such as MonoLoader for loading audio files are used.
Included in this repository are a Python Notebook with the code for the features extraction (AudioFeaturesBasedPlaylist.ipynb), the result of the analysis of the 2100 MusAV audio files (processed.csv) and a Streamlit interface to generate the playlists (Playlist_streamlit.py).