Videos uploaded to social media platforms are scanned against a database of copyrighted songs. If the Content ID is matched, the copyright holder can either claim money from the video or have it deleted.
How can we generate legal audio for video makers on social media platforms?
What it does
We start with an audio file we want to legitimify. This song structure is used as a template to spectrally recreate a generative cover from our database of small pieces of royalty free audio.
How we built it
1) created a large database of royalty free music
2) extracted audio fingerprints with 40 MIR features (pitch, timbre, etc) using Librosa
3) analyze a song to recreate
4) search all the audio for similar content in our database using K-d trees with SKlearn
5) rebuild the original song by concatenating all of the similar sounding audio from the database