Content based filtering methods like the NLP and audio analysis approach can be used to diversify the results that you'd get from collaborative filtering (or add 'noise' to recommendations, as someone up the thread wished for).

I'm guessing they evaluate in many ways. E.g. play counts for online evaluations, accuracy compared to some ground truth for offline evaluations

I really don't understand why collaborative filtering isn't the sole approach. It's basically exactly what you want if you have that volume of data.

Relying on collaborative filtering (using play counts, as Spotify do) can lead to non-diverse and non-serendipitous recommendation. If I like Abbey Road then a typical collaborative filtering scheme might tell me I like Sgt Pepper's and Let it Be. Excellent recommendation, but useless if I'm already familiar with the band and the genre.

There might be an unknown Japanese band that'd suit my tastes perfectly and I'd never know with Collab filtering.

That just seems like a narrow net problem than an issue with the approach. Assuming I have some distance function that tells me who's "similar" to me and it's malleable, I should be able to reach those outliers.

Yeah, makes sense, and for sure works for many cases. But there's no meaningful distance between two sets of preferences that have a null intersection.

I've come across this just as a user with regard to film. I reckon that I've watched all of the 10/10 mainstream comedies that have been made in English. I want more 10/10 comedies, so I'll have to look at foreign languages films. But there isn't enough overlap of user preference, due to demographic separation I suppose, to get decent recommendation from other countries. If you start adding metainformation preferences (e.g. must not be in English, must be >9/10 stars etc) then you're back to Content-based recommendation territory (i.e. audio analysis as Spotify are doing)