My track Paris Thru the Tears scores a whopping 7.9! :-D
It was recorded on analogue synths (£1500 + £200), a HardSID slaved to an old P300 running Windows 98, an old Akai sampler (which cost £100), and 2 crappy mixing desks (£75 and £100) over the past month...
The only digital hardware was my Toshiba laptop running the Cubase sequencer.
But I have been making music since 1998.
[My point here is that getting a 7.9 doesn't mean huge studio expenses and watered-down content]
While I'm certain that the software works, it's obvious that it can only work to a limited extent. It has probably been fed a seed sample of past hits from which a set of general inferences have been inductively extrapolated. It then applies those inferred rules to score new songs fed into the algorithm. Nothing especially exciting about that.
Where it will always fail is in assessing 'novelty.' By definition, something that is novel is unexpected, new, innovated, or otherwise disruptive to our established norms. An algorithm based on norming songs to a generalized set of standards will, by definition, never understand the appeal of something novel.
Now, 99% of novel things are probably going to be bad. This doesn't set them apart from the 99% of derivative things that are also bad. But the 1% novel hits will be overlooked, missed, denied, or remixed if this algorithm is given actual decision-making weight. Just like teaching-to-the-test is frowned upon, production-for-the-algorithm will become the new conformity-inducing straightjacket for human creativity.
Finally, it's unclear what assumptions the algorithm creators are operating on, given that public tastes change over time. Are they operating solely in the pop music era created during the 1960s through to the present? Are they taking an even more restricted data set of pop music since the advent of the synth? How many genres of music do they include in their seed, and are the inferences drawn from them generalizable across genres? Across cultures to the pop charts of other countries?
Hopefully this type of algorithmic norming is less disastrous to the recording-promotion-distribution industry than it was to finance industry...
An internet community of musicians that I frequent collectively played with Hit Song Science during their "free trial" period.
Basically, we discovered that it doesn't do a very good job.
For example, I had a track that was 1:57 long with a very slow tempo and a minimal arrangement, however there was a slide guitar towards the end. This gave it a really high affinity with a bunch of US chart-topping country & western hits, and so it had a score of something like 8, which is just absurd.
So while I like the general principle, we discovered very quickly that its "similarity" metric is probably just too simple. There seemed to be a very weak correlation between songs that we thought sounded like hits, and the score they got on HSS.
6 comments
[ 3.1 ms ] story [ 23.8 ms ] threadIt was recorded on analogue synths (£1500 + £200), a HardSID slaved to an old P300 running Windows 98, an old Akai sampler (which cost £100), and 2 crappy mixing desks (£75 and £100) over the past month...
The only digital hardware was my Toshiba laptop running the Cubase sequencer.
But I have been making music since 1998.
[My point here is that getting a 7.9 doesn't mean huge studio expenses and watered-down content]
Where it will always fail is in assessing 'novelty.' By definition, something that is novel is unexpected, new, innovated, or otherwise disruptive to our established norms. An algorithm based on norming songs to a generalized set of standards will, by definition, never understand the appeal of something novel.
Now, 99% of novel things are probably going to be bad. This doesn't set them apart from the 99% of derivative things that are also bad. But the 1% novel hits will be overlooked, missed, denied, or remixed if this algorithm is given actual decision-making weight. Just like teaching-to-the-test is frowned upon, production-for-the-algorithm will become the new conformity-inducing straightjacket for human creativity.
Finally, it's unclear what assumptions the algorithm creators are operating on, given that public tastes change over time. Are they operating solely in the pop music era created during the 1960s through to the present? Are they taking an even more restricted data set of pop music since the advent of the synth? How many genres of music do they include in their seed, and are the inferences drawn from them generalizable across genres? Across cultures to the pop charts of other countries?
Hopefully this type of algorithmic norming is less disastrous to the recording-promotion-distribution industry than it was to finance industry...
Basically, we discovered that it doesn't do a very good job.
For example, I had a track that was 1:57 long with a very slow tempo and a minimal arrangement, however there was a slide guitar towards the end. This gave it a really high affinity with a bunch of US chart-topping country & western hits, and so it had a score of something like 8, which is just absurd.
So while I like the general principle, we discovered very quickly that its "similarity" metric is probably just too simple. There seemed to be a very weak correlation between songs that we thought sounded like hits, and the score they got on HSS.