https://essentia.upf.edu/models.html

Essentia allows us to extract "simple" music features (loudness, bpm, danceability) as well as more complex features through ML (pre-trained?) models (genre, mood, instrumentation, voiciness, valence, gender, acousticness, ...)

5Go / 8M+ sqlite dataset of spotify tracks and their audio features: https://www.kaggle.com/datasets/maltegrosse/8-m-spotify-tracks-genre-audio-features

We don't have to stick to Spotify's feature list (acousticness, danceability, energy, instrumentalness, liveness, speechiness, valence) as long as we find some vectors that allow for

• playlist creation based on human-understandable descriptions of musical feature space
• track matching based on similarity of features, that gives satisfyingly "similar" tracks
• varied enough vectors (i.e. 3 axes would be too little, 7 is more than enough)