I noticed a new audio_summary data point called Instrumentalness and took a quick look through the songs I have and I'm getting results that don't make sense. That is, they don't make sense if I assume that "instrumentalness" is a measure of how likely a song is to be an instrumental with no spoken words.
But I'm finding a mix with that thought. For instance Dave Brubeck's Take Five has an instrumentalness of 0.001 and Yes's Cans and Brahms has an instrumentalness of 0.943. Respectively, they have a speechiness of 0.037 and 0.044.
And then there's Argent's Hold Your Head Up - which is full of lyrics and definitely not an instrumental - which comes in with a speechiness of 0.023.
I have ~200 songs that I have tagged as Instrumentals, so I am 98% certain that every single one is indeed and instrumental. Checking them all, the highest Speechiness is 0.29, which makes me think that Speechiness is an excellent indicator. Instrumentalness is all over the map with these, ranging from 0.0001 to 0.974, and well distributed across the range. But checking other songs, I'm finding many non-instrumentals with very low Speechiness, like 0.023.
Can anyone shed any light on this? Or suggest a better way of identifying instrumentals?