Universal and Sony say their experts identified millions of label-controlled recordings inside Suno’s training data by fingerprinting audio files during discovery. The important part is what they actually tested. Experts were not prompting Suno, hearing something familiar, and calling the resemblance proof.
They worked with Suno’s training files in secure rooms at the office of Suno’s outside counsel. The process had two stages: fingerprinting each training file, then sending those fingerprints to Audible Magic for comparison with its reference database. Two separate trips were needed for the first stage, totaling two full weeks inside the secure environment.
Final match results arrived in January 2026 after information tied to non-plaintiff matches had been removed. Those results fed into the 60,202-recording Suno lawsuit, but fingerprinting was only the start of the legal sorting. A match can identify a recording while leaving ownership, licensing, registration, and liability to be checked separately.
Once controlled access began, experts could examine audio files actually present in the corpus made available under the protocol. Suno’s source code and training data were inspected in controlled rooms, while the fingerprinting procedure avoided handing the dataset to the labels.
The security fight mattered because the second stage required fingerprints from Suno’s data to be compared outside that room against Audible Magic’s content-recognition system. The parties spent months arguing over how that comparison would work before reaching an agreement. When the process finished, the plaintiffs said millions of Suno training files matched recordings supplied to Audible Magic by Universal or Sony affiliates.
For the Suno review, the labels said their experts generated fingerprints from the audio files inside the training dataset. Audible Magic then compared those fingerprints with its content-recognition database, which already contained recordings supplied by copyright holders. A returned match connected the tested training file with a known reference recording.
This distinction matters. Fingerprinting can survive ordinary changes that would make exact byte-for-byte hashing useless, such as a different encoding or compression format. So the result concerns the audio content, not a filename that happens to claim it is a particular song.
Suno was already familiar with the technology before this fight reached its current shape. In October 2024, it announced that Audible Magic would be integrated into user uploads to help identify unauthorized material. Platform moderation and litigation are different jobs, but the same underlying content-recognition idea sits at the center of both.
Sony described a manual process that compared ISRC numbers with registration certificates and removed works where its U.S. rights had clearly expired or otherwise ended. Universal described rights-confirmation work, including manual review and checks with artist representatives. The proposed amendment in the first case listed 61,026 recordings, while the separate complaint filed later asserts 60,202.
The split between fingerprinting and rights verification is easy to miss. A content match does not establish who owns every relevant right or whether a license existed. It also does not decide fair use, willfulness, or damages.
A fingerprint scan also does not reveal what a trained model learned from one particular file or whether the model can reproduce it. It addresses a narrower factual point: whether audio corresponding to a known reference recording was present among the training files that were tested. For a case built around disputed training inputs, that narrower point is still important.
The labels received their final match results on January 15, 2026, after non-plaintiff information was removed. Sony then compared matched works with registration and rights records, while Universal performed its own manual checks before choosing works to assert.
They worked with Suno’s training files in secure rooms at the office of Suno’s outside counsel. The process had two stages: fingerprinting each training file, then sending those fingerprints to Audible Magic for comparison with its reference database. Two separate trips were needed for the first stage, totaling two full weeks inside the secure environment.
Final match results arrived in January 2026 after information tied to non-plaintiff matches had been removed. Those results fed into the 60,202-recording Suno lawsuit, but fingerprinting was only the start of the legal sorting. A match can identify a recording while leaving ownership, licensing, registration, and liability to be checked separately.
The discovery process changed the evidence available
Before the labels got controlled access to Suno’s training data, their public case leaned heavily on generated outputs. They used carefully chosen prompts and pointed to melodic or vocal similarities as evidence that copyrighted recordings had been used during training. Useful, but indirect.Once controlled access began, experts could examine audio files actually present in the corpus made available under the protocol. Suno’s source code and training data were inspected in controlled rooms, while the fingerprinting procedure avoided handing the dataset to the labels.
The security fight mattered because the second stage required fingerprints from Suno’s data to be compared outside that room against Audible Magic’s content-recognition system. The parties spent months arguing over how that comparison would work before reaching an agreement. When the process finished, the plaintiffs said millions of Suno training files matched recordings supplied to Audible Magic by Universal or Sony affiliates.
Audio fingerprints turn recordings into searchable signatures
An audio fingerprint is a compact representation derived from the recording itself, not a title lookup, filename search, or ordinary file hash. Audible Magic’s compact perceptual audio signatures let unknown recordings be compared with signatures generated from known works. Matching happens at the content level rather than depending on whatever metadata happens to travel with a file.For the Suno review, the labels said their experts generated fingerprints from the audio files inside the training dataset. Audible Magic then compared those fingerprints with its content-recognition database, which already contained recordings supplied by copyright holders. A returned match connected the tested training file with a known reference recording.
This distinction matters. Fingerprinting can survive ordinary changes that would make exact byte-for-byte hashing useless, such as a different encoding or compression format. So the result concerns the audio content, not a filename that happens to claim it is a particular song.
Suno was already familiar with the technology before this fight reached its current shape. In October 2024, it announced that Audible Magic would be integrated into user uploads to help identify unauthorized material. Platform moderation and litigation are different jobs, but the same underlying content-recognition idea sits at the center of both.
A fingerprint match is evidence, not the whole case
The match files carried more than a yes-or-no result. The earlier court filing says they included artist names, recording titles, ISRC numbers, and the entity that had delivered each matched recording to Audible Magic. Those fields gave the labels something they could cross-check against copyright registrations and their own rights records.Sony described a manual process that compared ISRC numbers with registration certificates and removed works where its U.S. rights had clearly expired or otherwise ended. Universal described rights-confirmation work, including manual review and checks with artist representatives. The proposed amendment in the first case listed 61,026 recordings, while the separate complaint filed later asserts 60,202.
The split between fingerprinting and rights verification is easy to miss. A content match does not establish who owns every relevant right or whether a license existed. It also does not decide fair use, willfulness, or damages.
A fingerprint scan also does not reveal what a trained model learned from one particular file or whether the model can reproduce it. It addresses a narrower factual point: whether audio corresponding to a known reference recording was present among the training files that were tested. For a case built around disputed training inputs, that narrower point is still important.
The labels received their final match results on January 15, 2026, after non-plaintiff information was removed. Sony then compared matched works with registration and rights records, while Universal performed its own manual checks before choosing works to assert.
