Sixteen tools that claim to tell you whether a song was generated. What each one actually costs, what it can and cannot see, and which numbers in the marketing are worth anything.
·12 min read
Suno and Udio put generation in reach of anybody with a text box, and the detection market that grew up around them now runs from free browser toys to enterprise contracts with no published price. Almost every vendor advertises accuracy somewhere north of 99%. Almost none of them say what they tested against.
That gap is the reason this page exists. Below is every AI music detector we could find that is actually reachable today, with pricing, coverage and access model taken from each vendor's own material and linked back to the source. Where a company publishes an accuracy figure we reproduce it and label it as their claim, not our finding.
On the scores: our own head-to-head measurements are still running, so every tool currently shows "not yet tested" rather than a number. We would rather show you nothing than a figure we cannot defend. The vendor claims and product details on this page are sourced and current.
All tools at a glance
AI music detection tools compared by category, free tier, pricing and vendor accuracy claim
Built for the case where the answer has to hold up in front of somebody who gets paid to argue with it. Reports separate what was measured from what was inferred, and every verdict carries its own false-positive rate rather than one headline number.
Pricing
Usage-based; report tier for legal work
Free tier
Free single-track check
Vendor claim
Published per-genre, with false-positive rate stated alongside
IRCAM has been the reference audio research lab in Paris for four decades, and that pedigree shows in the throughput: they advertise scanning 250,000 tracks in an hour, which is catalogue-sweep territory rather than one-song-at-a-time. The pricing opacity is the cost of entry.
The only entry on this list with a nine-figure evidence base behind it — Deezer says it tagged 13.4 million AI tracks across 2025, which is the closest thing the field has to real-world validation. The consumer tool scans playlists rather than files, so it answers a different question than a per-track detector does.
Pricing
Free for the playlist tool; licensing deals handled privately
Free tier
Yes — up to 100 playlists
Vendor claim
99.8% — roughly 2 misses per 1,000 AI tracks, under 1 false flag per 10,000 authentic songs (vendor claim)
Access
Free web tool, plus industry licensing since January 2026
The most transparent pricing page in the category, and the only vendor publishing a false-positive figure next to its accuracy figure — which is the number that actually matters when a wrong answer costs somebody their release. Twelve-model ensemble, sub-five-second responses.
Pricing
€12/mo (200 tracks) up to €2,399/mo (250k tracks); custom from €9,999/mo
Free tier
14-day trial, 20 analyses; free checker on the site
Vendor claim
99.42% accuracy, false-positive rate under 0.6% (vendor claim)
Access
REST API with Python, Node and Java SDKs; webhooks
Analyses the full mix, the isolated vocal and the instrumental as three separate questions, which is the right architecture for hybrid tracks where a human wrote the beat and a model sang over it. Publishing no accuracy number at all is a conspicuous gap for a company that already runs audio ID at scale.
Aimed squarely at DSPs, distributors and collection societies rather than artists, with royalty misrouting as the problem being solved. The MCP server is a genuine first here — you can ask an agent whether a track is AI and get a structured answer back.
Pricing
Not published — demo request required
Free tier
Not disclosed
Vendor claim
Claims higher accuracy than competitors; no figure published
A moderation API that added music detection to an existing image and video stack, which makes it a sensible pick if you are already a customer and a strange one if you are not. The 12MB file ceiling rules out lossless masters.
Solves an adjacent problem: rather than asking whether a track is synthetic, TraceID decomposes it into vocal tone, melody and lyrics to find which protected element got borrowed. That makes it a licensing instrument more than a detector, and it belongs in the stack next to one rather than instead of one.
Scans finished masters for synthetic components and writes the result into metadata so it travels with the file downstream. The attribution-and-licensing framing is more developed than the detection claims, which remain unquantified in public.
Two decades of content identification infrastructure sitting inside platforms that already depend on it, now growing AI labelling on top. Distribution is the moat; the detection layer itself is newer and far less documented than the fingerprinting it rides on.
Grew out of a song-identification extension, so the detector is a feature on an existing habit rather than a destination. Fine for satisfying curiosity about one track; the absence of any published accuracy figure means it should not be the basis of an accusation.
Deserves credit for publishing a number under 90% when every competitor rounds up to 99, and for naming it as holdout accuracy rather than something vaguer. Accepting a Spotify URL instead of demanding a file removes the main reason people abandon these tools.
Priced for the working producer checking a handful of submissions a week, and the scan quota is sized to match. Returns a probability rather than a verdict, which is honest, but there is nothing published that tells you what any given probability is worth.
Built for voice rather than music, which matters because the failure mode it catches — a cloned vocal over a real instrumental — is the one that most music detectors miss entirely. Run it alongside a music detector, not instead of one.
Pricing
Enterprise — not published
Free tier
Demo available
Vendor claim
Strong published cross-generator figures
Access
API and real-time deployment
Detects
Synthetic and cloned speech across many voice models
The training corpus is the widest published in voice detection, and the fraud-and-call-centre lineage means it was hardened against adversaries who actively wanted to beat it. Music is out of scope, so it answers the vocal question and nothing else.
Pricing
Enterprise — not published
Free tier
Demo available
Vendor claim
Trained on 350+ generation tools, 20M+ utterances, 40+ languages
A first-party classifier that is close to authoritative on its own generator and close to useless on anybody else’s. Treat a positive as near-conclusive and a negative as telling you nothing beyond "not ElevenLabs".
Vendor accuracy claims in this category cluster suspiciously tightly around 99%, and almost none of them state the test set. A detector that never flags anything scores 100% on human music and 0% on AI; a detector that flags everything inverts it. One number cannot describe both halves, so we report both.
AI recall
The share of known AI-generated tracks a tool correctly flags. Sources span Suno, Udio, MusicGen, Stable Audio, Mureka and Riffusion, because a detector tuned on Suno alone falls over the moment it meets anything else.
Human pass rate
The share of verified human recordings a tool correctly leaves alone. This is the number that decides whether a detector is safe to point at a real catalogue, and it is the number vendors are quietest about.
Hybrid handling
Real infringement rarely arrives as a clean fully-generated file. An AI vocal over a human instrumental defeats most whole-track classifiers, so we score partial cases separately instead of folding them into an average that hides them.
Stated limits
Whether the tool tells you what it cannot do. A detector that abstains when the evidence is thin is more useful in a dispute than one that always produces a confident answer.
Caveats
Free tiers are tested where a paid tier was not made available to us, and the paid model may perform differently.
Vendor-published figures are reproduced as claims, attributed, and never merged with our own measurements.
Detection performance moves whenever a generator ships a new version, so every result carries the date it was measured.
Before you trust any of them
Four things worth knowing whichever tool you land on, learned mostly from watching people act on a number they had no way to interpret.
01
Read the false-positive rate, not the accuracy
A tool advertising 99% accuracy on a test set that is 99% AI has learned to say yes. The number that protects you is how often it flags human work, and a vendor who will not publish it has usually decided you should not see it.
02
Test on your own catalogue before you buy
Every detector on this page was trained on somebody else's distribution of music. Run twenty tracks you already know the provenance of — including the ones with heavy production — and see what comes back.
03
One detector is a signal, not a verdict
Independent tools fail on different material. Agreement between two of them is worth far more than a high score from either alone, which is why serious disputes should never rest on a single classifier.
04
A score is not evidence
If the outcome is a takedown, a royalty hold or a legal claim, somebody will eventually ask how the number was produced. Screenshots of a percentage do not survive that question; a documented method with a stated error rate does.
Detection you can put in front of a lawyer
Most tools on this page return a percentage. When the outcome is a takedown, a royalty hold or a claim, somebody will ask how that percentage was produced. Our reports answer that question in writing.