To your first issue, there are three ways I can think of that music recommendations are usually built: - collaborative filtering: i.e. People who listened to x tend to listen to y, so if you like x we will recommend y - metadata: x and y share multiple genre tags, or have contributed to releases in common, or release music in the same time period, etc - audio data: x and y have similar tempo, prominent instrument timbres, absence or presence of vocals, time signature, major or minor-ness, key

However, I've never heard of lyrics being used as data or metadata for determining similarity of tracks.

For one thing, it's not necessarily a common use case. I did college radio and we would do themed shows where all the songs are about food, for example, but 99% of the time of I listen to a song that happens to mention ice cream, I don't want/need the next song to mention ice cream.

On top of that, it's not obvious what type of lyrical similarity is desired. Do you want to match lyrical sentiment? (Happy songs with happy songs, whether they're about girls or cars or cooking) Or theme (relationship songs, positive or negative) or words in common or phrases in common.

It's definitely an interesting idea that I'd love to see toyed with, but I'm not surprised that its not, given how much effort it might be for an unclear and possibly uncommonly desires result. Not to mention that Spotify usually doesn't have a canonical source for lyrics of a given track anyway.