Project seeking to bridge the semantic gap between music audio and user preferences
The automated curation of music playlists has become a signifcant problem in the last decade, with the colossal rise of streaming platforms e.g. Spotify.
Current state-of-the-art recommender systems depend on the collaborative filtering model. Nevertheless, these systems experience the cold start problem; they break down when no historical data is available and as a result they cannot recommend new and unpopular songs.
In this project, the author proposes a new recommendation model that uses music audio and deep learning to produce playlists based on a given query song.
details explained here: https://medium.com/@silvercloud438/how-i-taught-a-neural-network-to-understand-similarities-in-music-audio-d4fca54c1aed
to use this system:
-
download a dataset (FMA,GTZAN)
-
feature convert data to mel-spectrograms using featureConverter script
-
change file path in model script to feed data and train CNN
-
keep an ordered list of songTitles in text file for library of songs
-
run CNN in predict script
-
run playlist generator
-
output playlist :)
have fun!