Interesting approach but a lot of songs I would consider falling into this category score very low:
Here are a few examples from the original reddit post:
The Velvet Underground - Heroin (2.95)
The Beatles - A Day In The Life (0.00)
Other songs with low correlation:
Lynyrd Skynyrd - Free Bird (0.38)
Metallica - Hero Of The Day (3.72), One(3.47)
I think the reason is that tempo and compression (which I associate with tension) are not taken into account but play a big part in our hearing experience. This data may be harder to extract, though. Anyway interesting work!
I find this truly amazing, not because of the technologies involved, but mostly because I find it fascinating that we can now easily process a library of over a million song and find which have a “slow build”.
This is very cool - one quibble though, there doesn't seem to be any test of 'slow' in the slow build. You could add a linear regression slope and select for songs where Pearson coefficient is high, the slope is small and the last decile volume/first decile volume is high is large.