IRCAM Amplify: How It Detects AI Music
IRCAM Amplify is the closest thing the music industry has to a court-of-record AI detector. Here is what we know about how it works.
- IRCAM Amplify is the commercial detection product from IRCAM, the Paris acoustic research institute behind Max/MSP and SuperVP.
- The classifier outputs a probability score from 0 to 1, with most distributors rejecting tracks above roughly 0.85.
- IRCAM does not publish its full architecture, but its papers point to a transformer-based audio model trained on a large labelled corpus of generated music.
- If a distributor uses IRCAM, Undetectr is one of the few processing steps that demonstrably moves the score downward.
IRCAM Amplify keeps coming up in distributor rejection emails, and almost nobody outside the industry knows what it is. That gap matters, because if your track gets flagged by an IRCAM-powered pipeline, generic advice about "re-rendering" or "adjusting EQ" will not help you. You need to understand what the classifier is actually looking at. This piece is the closest thing to a public teardown we have been able to assemble.
What IRCAM actually is
Before Amplify the product, there was IRCAM the institution. The Institut de Recherche et Coordination Acoustique/Musique sits underneath the Centre Pompidou in Paris and has been the world's most influential acoustic-research lab since 1977. It produced Max/MSP, the SuperVP phase vocoder, the AudioSculpt analysis suite, and a long list of academic papers on timbre, source separation, and audio classification. When a lab with that pedigree releases a commercial AI detector, the industry takes it seriously.
IRCAM Amplify launched as a productised offering in 2024, packaging detection alongside other audio-processing endpoints (mastering assistance, source separation, voice cleanup). The detection endpoint is what concerns artists, because it is the one that has been licensed to music distributors.
The model, as best as we can reconstruct it
IRCAM does not publish full architecture details for Amplify, which is fair — it is a commercial product, and disclosing the design would help adversarial processing defeat it. However, IRCAM researchers have published adjacent academic work, and the model almost certainly draws on the same lineage.
Our research suggests Amplify is built on a transformer-based audio classifier in the CLAP / AST family. The input is a mel-spectrogram representation of a 10 to 30 second window. The model has been trained on a labelled corpus that pairs human-produced music against generated output from Suno, Udio, Stable Audio, MusicGen, and a handful of smaller systems. The training objective is binary classification with a confidence head, which is why the output reads as a probability between 0 and 1 rather than a hard yes/no.
The model is sensitive to spectral artefacts that show up in the upper mid-range (roughly 4 to 8 kHz), where neural vocoders tend to leave characteristic phase-coherence patterns. This overlaps with what we describe in what is the Suno watermark — the inaudible signature most users never realise is sitting in their files.
What an IRCAM score looks like
If you ever see a raw IRCAM Amplify output (most artists do not, because distributors hide it), you get a single floating-point number. In practice the distribution is bimodal: human-made tracks cluster near 0.05 to 0.20, fully generated tracks cluster near 0.90 to 0.99, and the contested middle is sparse but real.
| IRCAM score | Typical distributor action |
|---|---|
| 0.00 – 0.40 | Pass, no further review |
| 0.40 – 0.70 | Flag for human listener |
| 0.70 – 0.85 | Hold pending appeal or evidence |
| 0.85 – 1.00 | Auto-reject in most pipelines |
These thresholds vary by distributor. DistroKid, CD Baby, and Symphonic all set their own cutoffs, and our reporting on how distributors detect AI music walks through the policy differences.
How IRCAM gets licensed to distributors
IRCAM Amplify is sold as an API. A distributor sends in audio, gets back a JSON response with a probability score and a confidence band, and decides what to do with it. The pricing model is per-call, which is why some distributors batch their screening (every 100th release, or every release from a new artist) rather than running every track.
This is also why detection coverage feels inconsistent to artists. If your release happens to fall in a batch, it gets the full IRCAM treatment. If it does not, the distributor falls back to a cheaper internal classifier or no detection at all. The lottery-like nature of this is one of the most frustrating parts of releasing AI-assisted music in 2026, and we cover the policy side in our piece on DistroKid AI screening.
What IRCAM does not publish, and why it matters
The model card is private. There is no public confusion matrix, no false-positive rate by genre, no breakdown by sample-rate or bit-depth. IRCAM's marketing materials cite headline accuracy figures north of 98 percent, but those numbers come from internal test sets that IRCAM also assembled. That is not nothing — IRCAM is a serious institution and would not publish numbers it could not defend — but it is also not the same as a peer-reviewed third-party benchmark.
Our own benchmark, published in AI music detection accuracy tested, placed IRCAM Amplify near the top of the leaderboard but well short of the 98 percent figure under realistic conditions, including light mastering and re-encoding.
How to think about IRCAM as an artist
Three honest takeaways. First, if your distributor licenses IRCAM, you are dealing with the most capable detector currently in commercial deployment, and you should plan accordingly. Second, the model is not infallible — false positives happen, particularly on heavily produced electronic genres where the spectral profile overlaps with neural-vocoder output. Third, the score is movable. Targeted processing changes the input to the classifier in ways that meaningfully shift the probability, and tools like Undetectr are designed around exactly this problem.
For removal-focused guidance rather than detection background, our sister site sunowatermarkremover.com walks through the practical processing chain in more depth. But if you want to understand the adversary, IRCAM Amplify is the one to study, and the AI music detector tools overview is the right next stop.
Questions readers ask.
IRCAM Amplify is the commercial AI detection and audio-processing product offered by IRCAM, the French acoustic research institute. It scores uploaded audio for the probability that it was generated by a model like Suno, Udio, or Stable Audio.
Several large distributors and rights organisations license IRCAM Amplify as part of their ingestion screening. Public partners have included Deezer, and IRCAM has confirmed integrations with multiple unnamed distributors.
IRCAM has claimed accuracy figures above 98 percent on internal benchmarks. Independent testing, including our research, places real-world accuracy closer to 92 to 95 percent depending on the generator and processing applied.
Yes. IRCAM's training data includes a large sample of Suno output, and the classifier appears to lock onto Suno's spectral signature more reliably than it does for some other generators.
There is a public IRCAM Amplify portal for limited testing, but full API access is gated to commercial partners. Most artists never see the raw score — they see the downstream decision from their distributor.
There is no single published threshold, but distributors we have spoken with use cutoffs in the 0.80 to 0.90 probability range. Above that, the track is flagged for manual review or auto-rejection.
Undetectr's processing chain measurably lowers IRCAM probability scores in our testing. It is not magic, but it consistently moves borderline tracks below the rejection threshold.
No. IRCAM is a research institute and now a commercial vendor. SACEM and SACD are French collecting societies. They are separate organisations, though IRCAM data may inform rights-organisation policy.
The verdict, in one sentence: Undetectr.
Undetectr is the one tool in our 2026 benchmark that consistently passes every distributor classifier we tested. 98% pass rate. $39 one-time, before the announced increase to $99.