Pre-launch · benchmark in design
Some records were made by people. That ought to be worth stating.
Origin analysis for recorded music. A 0 to 100 score with the generative model named, the vocal and the accompaniment scored separately, and a dated finding you can put in front of a distributor, a curator or a tribunal.
The detector is not live yet. We are publishing the method and the benchmark design before the product, which is the opposite of how this category normally works.
What is actually happening
Ninety thousand fakes a day
Nobody uploading those tracks is required to say what made them. A playlist curator cannot tell. A competition judge cannot tell. An A&R listening to two hundred submissions a week cannot tell. And an artist who spent four months in a room with a drummer is in the same royalty pool as a prompt.
Why we have not shipped a tool
A score you cannot audit is not evidence.
The detectors already on the market claim accuracy in the high nineties. When somebody finally checked one, the numbers did not hold.
A peer-reviewed audit published in Transactions of the International Society for Music Information Retrieval tested the one commercial detector that publishes a false-positive bound. Its advertised figure was under 1%. Measured against 30,000 tracks, the real rate was 4.7%. On a catalogue of 22 million recordings that is around a million human records wrongly called machine-made.
The same audit found the verdict flipped when a file was simply resampled to 22.05kHz, and that it stayed confident while being wrong, between 50% and 97%. The authors concluded detectors may be identifying encoder pipelines rather than authorship.
We would rather build the benchmark first and be measured against it than add another unaudited number to that pile.
What we commit to publishing
- False-positive and false-negative rates, separately, not a single blended accuracy figure
- Those rates broken out by generator, by genre, and by codec and sample rate
- The composition of the benchmark set, so the figure can be reproduced rather than believed
- Results under adversarial transforms: resampling, pitch shift, re-encoding, filtering
- The classes where we are weakest, printed next to the ones where we are strong
None of these numbers exist yet. That is the point of saying so here rather than inventing them.
The method
We take the song apart before we score it.
Most detectors score the finished mix. That is why a real singer over a generated beat sails straight through them.
01
Decode
Unpacked to raw audio so an MP3 of a real band is not punished for being an MP3.
02
Split
Vocal and backing separated and scored alone. This is where hybrids get caught.
03
Listen close
Breath, handling noise, micro-timing, and the periodic residue neural codecs leave behind.
04
Name it
SynthID and C2PA checked, then matched against the index to name the model and version.
Who this is for
Four roles, one question, four standards of proof.
Artist or producer
A track beat mine on a playlist
Check it without an account. If it scores high and was never declared, you have something to say to the curator.
Label or A&R
Two hundred submissions a week
Batch the inbox. The stem split catches the ones where a real singer is fronting a generated bed, the case that gets past everything else.
Distributor or DSP
Undeclared AI in the delivery queue
Score every delivery before acceptance. The declared-versus-detected gap is the part a supplier agreement can actually enforce.
Sync or contest
The brief said original composition
A dated finding per entry, defensible if a disqualified entrant argues, including the threshold band and its error rate.
The tool
Drop it in. Find out.
When it opens, three checks a day will be free with no account and no card, permanently. Your file will be analysed in memory and discarded the second the scan ends.
The verdict is never the paid part. Paid plans buy the stem split, the watermark check, the paperwork and the volume.
- MP3, WAV, FLAC or M4A, or a streaming link
- A 0 to 100 score for the vocal and the backing separately
- The generative model named and versioned where it can be named
- Audio never stored, never sold, never used to train anything
Join the waitlist
We will write when there is something real to show.
No drip sequence, no launch countdown. One email when the benchmark is published, one when the tool opens.
If you are a label, a distributor or a platform, say so and say roughly what volume. We will come to you first.
Common questions
Questions people actually ask
That is what we are building. The honest answer today is that the whole field is less reliable than its marketing suggests, which is why we are publishing our method and our error rates before we publish a product. You will get a 0 to 100 score rather than a yes or no, because a score can be audited and a verdict cannot.
That is the case we are building for. Scoring the finished mix averages the answer away: a human vocal over a generated bed reads as ambiguous and sails through. Vocal and accompaniment have to be separated and scored on their own.
No. The design commitment is that files are analysed in memory and discarded when the scan completes. Never stored, never sold, never used to train anything, on free and paid plans alike.
Because shipping a detector that calls real human records machine-made would be worse than shipping nothing. The published research shows the only independently audited commercial detector missing its own stated false-positive rate by roughly five times. We would rather build the benchmark first and be judged against it.
The verdict will be free, permanently, with no account. Paid plans buy the breakdown, the paperwork and the volume. They do not buy access to the truth.