DeepMind watermarked its own generated music, not music generally

What happened

On 16 November 2023, Google DeepMind announced that it had extended SynthID, its watermarking tool first built for images, to music generated by its Lyria model. The announcement ties the watermark to Lyria's output and to the podcast-generation feature of NotebookLM, and frames it as a way for the company to mark its own generated audio for later identification, including by YouTube for its Dream Track experiment. The SynthID technology page, describing the tool as retrieved on 16 September 2026, confirms the same scope: it watermarks audio 'generated or published through' Lyria or NotebookLM's podcast feature, not audio in general.

What the documents say

Both documents describe the watermark as embedded by converting audio to a spectrogram and altering it in a way stated to be 'inaudible to the human ear' and not to 'compromise the listening experience'. On robustness, the announcement states the mark is designed to survive 'noise additions, MP3 compression, or speeding up and slowing down the track', and that SynthID can identify whether only part of a song came from Lyria. Neither document gives a numeric detection accuracy or describes testing against adversarial removal, cropping, pitch-shifting beyond simple speed change, or re-recording through a microphone.

Why it matters for makers

The mechanism is embedding, not scanning: SynthID marks audio a compliant Google tool generates, it does not scan arbitrary files to say whether some other generator made them. A track that sounds Lyria-like but was never run through Lyria will not carry this watermark, and its absence is not evidence of anything about origin. For a producer asked to prove a stem is not AI-generated, or asked to check whether a submitted vocal is watermarked, the practical scope is narrow: this specific mark only ever applies to Google's own generation pipeline, described as it stands on the date these pages were retrieved.

What to check before you use it

If a collaborator supplies audio and claims a SynthID check clears it, confirm which detector was used and whether it can only recognise Google's own watermark rather than watermarks or fingerprints from other systems. Do not treat an unmarked file as proof of human authorship, since most generators in use, per the documents seen, are not stated to embed SynthID at all. This is an editorial caution beyond the announcement itself: robustness claims describe specific transformations tested by the developer, not an exhaustive list of what a determined actor could do to strip a mark.

• Was this specific audio generated through a pipeline that the developer states embeds SynthID, or another tool entirely?
• Which transformations has the developer stated the watermark survives, and does your intended use fall outside that list?
• Does the absence of a SynthID mark tell you anything at all about whether other generators were involved?

A watermark answers one narrow question about one company's own generated audio. It does not function as a general test for AI involvement in a track, and the source documents do not claim that it does.

Continue in Rights & release, or A provenance standard records edits, not the truth of a file.