Measured, not asserted
Each experiment states what was tested, how, and what the numbers were — including when the result was boring.
Research
Things we measured while building this product, published with the method and the limitations so you can check them. Short list by design — a page goes here only when there is an experiment or a dataset behind it.
2026-08-10
Measuring centre-channel vocal separation
Band-limited mid/side cancellation removed a centre-panned test tone by 29.5 dB — about 97% of its amplitude — while a hard-panned tone in the same file lost only 0.5 dB. That gap is the whole basis of karaoke-style vocal removal, and it is measurable rather than a matter of opinion.
2026-08-10
What AI music sitemaps reveal about the market
Public sitemaps show where a company has genuinely invested. Suno publishes three separate programmatic surfaces totalling roughly 1,090 URLs, of which about 967 are genre and style pages — and about 16 of its 69 tool landing pages target vocal removal alone.
What counts as research here
A page appears in this section only if there is an experiment or a dataset behind it. That rules out most of what gets published as research in this industry: a rewritten press release, a survey with an undisclosed sample, or a benchmark with no method attached. If we cannot tell you how to reproduce a number, we do not publish the number.
Each page states its method in enough detail to repeat, reports the measurements as they came out, and lists the limitations as prominently as the results. The limitations section is not a disclaimer — it is the part that tells you whether the finding applies to your case. A measurement on synthetic test tones says something precise about an algorithm and much less about a commercial master, and saying so is the difference between research and marketing.
Figures carry the date they were captured. In a field where model versions ship monthly, an undated benchmark is closer to a rumour than a fact, and anything here older than a quarter should be re-checked before you rely on it.
Inside
What we measured
Experiments run in-house, with the method published so you can repeat them.
1 / 5
Most claims in this field are unsourced
Publishing the method is the cheapest way to be trustworthy in a category where almost nobody is.
Negative results are useful
Experiments that found nothing are published too, because knowing what does not work saves time.
It keeps our own copy honest
A measurement we published is a claim we cannot quietly inflate later.
Readers can check us
The best correction is someone re-running the test and telling us we were wrong.
Frequently asked questions
What kind of research is published here?
Original measurements — how separation quality is tested, how search engines index AI music sites — with the method published alongside every number so the test can be repeated.
Is this research peer reviewed?
No. It is in-house measurement published with its method and its limitations stated, which is a different thing from academic publication and should be treated as such.
Can I cite these studies?
Yes, with the date and the stated limitations. Each study carries what it did not test, which matters as much as what it found.
Why publish the limitations?
Because a measurement without its limits is a marketing claim. Anyone repeating the test needs to know where it stops being valid.
Do you publish negative results?
Yes. A method that failed to show what we expected is more useful to a reader than one that confirmed it, and omitting those is how a research page becomes marketing.
Can I request a study?
Suggestions are welcome, particularly where a claim is circulating with no measurement behind it. Those are the most useful things to test.