a row of spikes nobody trained in

there is a sentence at the top of this paper that the researchers did not write. it belongs to the chief executive of suno, and they put it there on purpose.

i would rather talk about what four people at deezer did after this quote, because generated music turns out to leave a mark, and the mark has nothing to do with taste and the usual debate on ai.

the frame

evenly spaced peaks came from a process. musically spaced peaks came from the music.

where the spikes come from

the models build audio by making small things bigger. the layer that does it inserts zeros between the numbers it already has, then runs a short filter across the result. stretch, then filter.

inserting zeros makes copies of a spectrum. that is not a side effect, it is what the arithmetic does. these models also carry a spike at zero hertz, because the layers add a constant and clip anything negative. copy the spectrum, and you copy that spike with it. one becomes a row of them, evenly spaced, all the way up.

in the audio, that row is the hiss you hear.

here is the most important part for audionerds. stretch, then filter is the same operation your oversampling plugins run. every time a saturator or a clipper oversamples, it stuffs zeros between your samples and filters the result. a good one filters hard and you never hear it. a cheap one leaves you the copies.

so this is not a story about robots. it is a story about ordinary arithmetic, and what it leaves behind when nobody cleans up after it.

what to do with that

when you see a comb of equally spaced peaks in an analyzer, stop trying to EQ them.

they are not notes. music is built on ratios, not on addition, so nothing a performer plays lays itself out at a fixed hertz spacing all the way up. not a piano, not a room, not a voice. fixed spacing is a signature of your signal path.

which means it is a source problem. change the plugin, turn its oversampling up, or re-render the stem. notching them one at a time is a losing afternoon, and you will dull everything around them for nothing.

this is most of why i spent so long on antialiasing in WARM instead of on presets. the interesting part of a saturator is not the curve. it is what the curve leaves behind.

the other tell, and this one is mine

the paper is about the spikes. but every time i have put generated audio on an analyzer, there is a second thing: the top end just stops. somewhere around 15 kHz there is a shelf, and above it almost nothing.

that is what a lossy codec does. so my read is that a lot of this was trained on mp3s, and the models learned a ceiling that was never in the music.

that is my observation and not the paper’s, and a missing top octave can come from more than one place. but it costs nothing to look.

and it works the same way round as the spikes: that shelf is absence, not a dip. there is nothing above it to lift, so reaching for an exciter is the same losing afternoon.

since it will come up: no, i do not make music this way. i find the mechanism interesting, which is a different thing.

the number at the back of the paper

the fingerprint is structural - they trained the same architecture four separate times and the spikes landed in the same places every time, so the detector does not need to be clever. on the models it was trained against it scores 100.00.

then they show it one it had never seen, and it catches about four in ten. real music in that same table is still caught at 99.97, so nothing fell over. the authors think the architecture changed between versions - of course it gets harder to notice over time but you can definitely still hear it.

none of the numbers in the film are mine, and i say so on camera. the paper is open: arxiv.org/abs/2506.19108.