AI Music Detectors, Compared

An AI music detector estimates whether a track was created by a model like Suno or Udio. Most detectors look for production artefacts in the audio itself rather than relying on a hidden watermark.

The two tools most often compared in the music industry are IRCAM Amplify, developed by a French sound research institute, and ACRCloud, an audio recognition company. Both are designed for distributors and streaming services, not individual listeners.

Audio provenance doesn't work the same way as it does for images or text, and that difference determines which tool will be useful to you. This page breaks down the three signals involved, compares the leading detectors, and explains the part tied to revenue: what happens when a track reaches a distributor.

IRCAM Amplify AI Music Detector vs. ACRCloud AI Music Detector: Side-by-Side

The Three Signals an AI Music Detector Can Read

Audio provenance comes from three separate signals, and each one behaves differently. Confusing them is the reason most detector roundups give bad advice.

The first is an embedded audio watermark: an inaudible pattern the generator writes into the waveform itself. Google DeepMind says the SynthID audio watermark cannot be altered by common modifications like adding noise, MP3 compression, or changing the speed of a track.

The second is file-level provenance, such as C2PA Content Credentials attached to the file. That is the weakest link, because re-encoding or a format conversion strips it, as our C2PA guide explains.

The third is statistical inference from production artefacts. That is what almost every commercial AI music detector actually does, and it is a guess rather than a reading.

• Embedded audio watermark: written into the waveform; can survive compression and format changes.
• File metadata or C2PA: easy to strip, so absence proves nothing.
• Production artefacts: a statistical estimate, with false positives in both directions.
• Vocal cloning checks: a separate question about whose voice it is, not whether a model made the track.

IRCAM Amplify vs. ACRCloud

IRCAM Amplify and ACRCloud both classify audio without needing a watermark, and both sell to the same buyers. They differ in what the result tells you and how you can try them.

IRCAM Amplify markets throughput. Its product page claims 99% accuracy with less than 1% false positives, and says the system can scan more than 250,000 tracks within an hour depending on server capacity. There is no public price and no self-serve free tier; access starts with a request.

ACRCloud analyses a track in parts. It reports the full track, the isolated vocals, and the accompaniment separately, and names the likely generator among platforms including Suno, Udio, ElevenLabs, Seed Music, MiniMax, Mureka and Riffusion. It offers a 14-day free trial with full API access and no credit card.

Treat the accuracy gap as a marketing gap, not a measured one. IRCAM Amplify publishes a headline figure; ACRCloud publishes none. Neither number has been independently audited. An Associated Press reporter who uploaded a handful of self-generated tracks to IRCAM Amplify in July 2025 got AI probabilities of 81.8% to 98% and a correct attribution to Suno, while old MP3s scored low, but AP cautioned that a handful of results is not a sign of overall accuracy and noted that other detectors tried for the same story returned results that were either inconclusive or flagged AI songs as human-made and vice versa.

• IRCAM Amplify: catalogue-scale scanning, vendor-claimed 99% accuracy, quote-based, no free tier.
• ACRCloud: component-level results, names the generator, 14-day free API trial.
• Both: work on audio alone, so a stripped file is still analysable.
• Neither: publishes an independently audited false-positive rate; press spot checks are too small to serve as one.

Sightengine, AHA Music, Deezer, and Others

Several other AI song detectors cover the cases IRCAM Amplify and ACRCloud do not, including the free check most listeners want.

Sightengine returns a confidence percentage and states plainly that it works even where metadata has been stripped and no watermark is available. It runs on a free account with monthly credits, then paid tiers.

AHA Music is the free front door most people land on. It needs no signup, allows a handful of checks per day, and is powered by ACRCloud, so the free consumer tool and the paid distributor tool share an engine.

Deezer built its own detector, deployed it in January 2025, and now licenses it to other companies. Google's SynthID Detector portal is a different kind of tool: it checks for Google's own watermark rather than guessing, and access has been limited to tester groups.

• Sightengine: free monthly credits, then subscription; explicitly watermark-independent.
• AHA Music: free, no signup, daily cap; ACRCloud engine underneath.
• Deezer: in-house detector, now sold to other platforms and rights bodies.
• SynthID Detector: reads a watermark, so it confirms Google-model audio and nothing else.

Does Suno or Udio Watermark Its Output?

As of August 2026, Suno has announced audio watermarking but has not published the technical details, and Udio has not published a watermarking specification of its own. Verify both on the vendors' own pages before relying on either.

Suno said in a post dated 6 August 2026 that "in the coming weeks" it would adopt audio watermarking and fingerprinting technology so it can work with distribution platforms against fraud and misuse, and that it is screening uploaded audio and lyrics with Musixmatch and other third-party providers. Reporting on the announcement identified the Musixmatch component as that company's Sentinel copyright-detection service, and quoted chief executive Mikey Shulman saying the tools are "designed to be durable and resistant to tampering, without affecting the listening experience".

What Suno did not say is as important: the timing is given only as "the coming weeks" rather than a dated launch, there is no confirmation of whether it uses SynthID or a scheme of its own, no published technical detail, and no statement about tracks generated before the change. Suno's separate 10 August 2026 post covers download caps and states that trial downloads are for personal use only, with commercial rights applying to songs downloaded on paid plans. Our Suno pricing guide has the tier detail.

Udio settled with Universal Music Group in October 2025 and is building a licensed platform where creations stay inside the service rather than being downloaded and distributed. A walled garden is a provenance strategy, but it is not a watermark, and it says nothing about files already exported from the older product.

• Suno: watermarking and fingerprinting announced 6 August 2026 for "the coming weeks"; implementation details not public at the time of writing.
• Udio: licensed platform with no off-platform downloads; no published watermark spec confirmed.
• Neither: should be assumed to carry a readable mark on an older exported file.
• Where to check: the Suno blog and Udio, plus our AI watermarking guide for the wider vendor roster.

How to Tell If a Song Is AI Generated

Work in three passes, and run them in this order: provenance first, detector second, listening tells third. Reversing the order is how people talk themselves into a wrong answer.

Start with provenance. Check the credits, the songwriter and publisher fields, the release history of the account, and any Content Credentials attached to the file. A track with no named writer, no publisher, and an account whose entire catalogue appeared last week tells you a lot before you press play.

Then run two detectors, not one. Different engines are trained on different data, so agreement between two tools is worth far more than a single high score. Where the model family is Google's, the SynthID explainer covers the watermark check that beats any statistical guess.

Listening tells come last, because they are the least reliable and the easiest to fake. The concrete ones: a faint hiss layered over the vocal that gets louder in quiet passages, breaths or consonants landing where no singer would put them, an instrument cutting off mid-phrase, and choruses that repeat identically instead of building. Lyrics are the other giveaway, hitting clean rhymes with no names, places, or specific images in them.

• Pass 1: credits, publisher, upload history, Content Credentials.
• Pass 2: two independent detectors, and treat disagreement as unresolved.
• Pass 3: vocal hiss, misplaced breaths, identical repeats, lyrics with no specifics.
• Never: accuse anyone on a single detector score.

What Happens When AI Music Reaches a Distributor

Detection stopped being an academic question when streaming services started screening uploads and withholding royalties. That is where the real money sits, and it is the part generic detector roundups skip.

Deezer reported in July 2026 that fully AI-generated tracks passed 50% of daily new-music uploads, around 90,000 tracks a day, while accounting for only 1% to 3% of actual streams. Deezer says it demonetises fraudulent streams, and that up to 85% of streams on fully AI-generated tracks were fraudulent in 2025. Detected tracks are labelled for listeners and excluded from algorithmic recommendations and editorial playlists.

Spotify took the disclosure route instead. It backs a DDEX industry standard that lets creators declare AI's role in vocals, instrumentation, or post-production, shown in song credits, and it stated that the aim is not punishing artists who use AI responsibly or down-ranking tracks for disclosing how they were made. That disclosure now runs through distributor upload flows, starting with DistroKid.

For an independent artist, the asymmetry is brutal. A wrongly flagged track can lose playlist placement, algorithmic reach, and royalty income at once, and the appeal runs through your distributor rather than a button in an app. Disclose accurately at upload, keep session files and project exports, and check the platform's creator support pages for the current dispute route before you need it.

• Deezer: labels fully AI tracks, drops them from recommendations and editorial, demonetises fraud.
• Spotify: disclosure via DDEX credits, plus a spam filter and an impersonation policy.
• Distributors: now collect AI-role declarations at upload; false declarations create their own risk.
• Your defence: dated project files, stems, and session exports, kept before a dispute starts.

Why Detectors Misfire on Produced Human Recordings

Statistical detectors flag human recordings when the production sounds like the thing they were trained to catch. Certain human music is genuinely hard to distinguish from generated audio on artefacts alone.

Quantised electronic music is the clearest case. Grid-locked drums, synth-heavy textures, and near-perfect timing produce the same low-variation profile a generator does, because a human producer removed the variation on purpose.

Heavy mastering does the same at the other end. Loudness maximisation flattens dynamics, so a modern pop master and a generated track can share a suspiciously similar dynamic fingerprint.

Hybrid workflows are the messiest. A songwriter who writes the lyrics and melody but uses a model to render the arrangement sits genuinely between categories, and a binary AI-or-human label describes that track badly no matter which way it lands.

• Highest risk genres: EDM, synthwave, techno, and anything tightly quantised.
• Second risk: heavily limited, loudness-maximised pop masters.
• Third risk: part-AI hybrid production, which no binary label describes correctly.
• Lowest risk: live-tracked acoustic recordings with natural timing variation.

Which AI Music Detector Should You Use?

Pick by role, not by advertised accuracy. The right tool for a curious listener and the right tool for a distributor are not the same product.

A listener checking one song wants AHA Music or Sightengine's free tier, and should run both rather than trusting either. A label or catalogue owner scanning at volume wants IRCAM Amplify or ACRCloud on an API contract.

An artist worried about being mis-flagged should not be shopping for a detector at all. The useful work is accurate disclosure at upload, kept project files, and knowing the dispute route at each platform.

Verify current pricing on each vendor's own site. Every tool here either hides its price behind a sales conversation or changes tiers often enough that any figure quoted elsewhere goes stale.

• Listener: AHA Music plus Sightengine, free, cross-checked.
• Label or distributor: ACRCloud for platform attribution, IRCAM Amplify for throughput.
• Google-model audio: the SynthID Detector portal, which reads a mark instead of guessing.
• Artist: disclosure, archives, and dispute routes beat any detector subscription.

The Verdict

For catalogue-scale screening, IRCAM Amplify and ACRCloud are the two serious options, and they answer different questions. ACRCloud names the likely generator and separates vocals from accompaniment; IRCAM Amplify is built to sweep enormous catalogues quickly. Neither publishes an audited false-positive rate, so treat the marketing figures as claims.

For a single song, use a free checker such as AHA Music or Sightengine, and run both before believing either. Check credits and provenance first, because a stripped file with no writer and no history is more informative than any probability score.

The consequential shift is at the platforms rather than the tools. Deezer labels and demonetises, Spotify collects DDEX disclosures through distributors, and Suno has announced watermarking without publishing how it works. If you release music, accurate disclosure at upload and a dated archive of your project files protect you better than any detector you could buy.