Ircam Amplify AI Music Detector (AIMD)

The limits are stated in the API guide. "Supported formats: AIF, AIFF, FLAC, M4A, MP3, MP4, OGG, WAV, WEBM", "Track duration: 3 seconds to 30 minutes", "Maximum number of files per request: 20" and a "Recommended minimum MP3 bitrate: 256 kbps", because "lower-bitrate or highly compressed audio may affect detection performance" (docs.ircamamplify.io, Audio Input Methods, read 10 November 2026). The platform guide prints a shorter format list, "AIFF, FLAC, M4A, MP3, MP4, OGG, WAV", with "mono or stereo" channels and "20 files at a time" (docs.ircamamplify.io, Platform Guide, read 10 November 2026). Three input routes exist: upload to Ircam Amplify Storage, a presigned URL on the customer's own storage, and an ISRC code "for partners with ISRC processing enabled", 20 per request. The home page states throughput rather than a cap: "Scan more than 250.000 tracks within 1 hour, depending on server capacity" (ircamamplify.com, read 10 November 2026).

No price is published. The request page says "Our team will review your application and, upon approval, provide a demo and a secure onboarding link. As a primarily B2B-focused solution, our offering is tailored to professional use cases" and "Our pay-as-you-go plan starts from 2,000 tracks" (ircamamplify.com/request-aimd, read 10 November 2026). The platform guide describes two account kinds, Top-Up accounts with a track balance and a "Buy tracks" button, and Enterprise accounts with a contract period, an annual limit and a Team page; "Time Analysis is not available on Enterprise Core plans", which return the probability alone (docs.ircamamplify.io, Platform Guide, read 10 November 2026). The API error table names a free account with credits ("Not enough credits on a free account. Nothing is charged.") and states no allowance, so the free tier field records none published.

Retention is stated as deletion after analysis. Article 6.4 of the terms reads: "Audio files submitted to the Service are processed solely for the purpose of performing the requested analytical output operations and, unless otherwise expressly agreed in the Applicable Contract, are automatically deleted or rendered inaccessible after completion of the analysis", and the vendor "does not use the submitted audio content to create proprietary catalogues or exploit protected works" (app.ircamamplify.io, terms PDF, read 10 November 2026). The retention rank field takes that clause as a stated zero, the reading the index gave Paperpal and StealthWriter. The vendor keeps aggregated "Service Metrics" for research and benchmarking, and the AI Use & Transparency Framework says log retention periods "are defined in the Applicable Contract and/or security schedule" (app.ircamamplify.io, framework PDF, read 10 November 2026); that framework is what the site's Privacy Policy link serves.

The home page claim is a round number and the release note is a measured one. "99% [accuracy]" and "Less than 1% false positives" sit under the heading "The world's most accurate solution to detect AI-generated music" (ircamamplify.com, read 10 November 2026). The V2.3 note, for a release "scheduled October 22, 2026", states that "Performance is measured on Ircam Amplify's internal benchmark datasets" and prints, at the 50% threshold, "Accuracy 99.7%" ("Previously 99.6%") and a "False positive rate 0.14%" ("Previously 0.63%", a 78% relative reduction), and at the 80% threshold "Accuracy 99.1%" and "False positive rate 0.02%" (docs.ircamamplify.io, AIMD V2.3, read 10 November 2026). No dataset size, composition or date is published for the benchmark. The card writes the home page figure as 99% (accuracy) and quotes the release note beside it.

The vendor's own terms are more careful than its home page. Article 8.2 states that "no guarantee is given as to: the accuracy, completeness, or reliability of analytical outputs" and that outputs "may be affected by [...] transformations such as editing, clipping, transcoding, compression, remixing, stem separation, mastering, format conversion, excerpting, or other alterations", so "the Service does not guarantee definitive attribution, exhaustive detection, or error-free qualification of whether content is AI-generated, AI-manipulated, human-authored, or otherwise compliant with any legal or industry standard" (app.ircamamplify.io, terms PDF, read 10 November 2026). The framework adds that "AIMD outputs are decision-support information" and "do not constitute legal determinations, contractual determinations, or binding decisions".

The documentation sets no threshold. "There is no universal threshold for every use case. The appropriate threshold depends on the decision being supported and the relative impact of false positives and false negatives", and "Borderline results should be routed to human review rather than automatically accepted or rejected" (docs.ircamamplify.io, Platform Guide, read 10 November 2026). The results guide states that "The absence of an identified generator does not establish that a track is human-generated". A 99.7% on an unnamed internal benchmark and a terms clause that disclaims accuracy are both the vendor's statements, and the card keeps each with its source.