Google introduces its Lyria 3.5 music model to all Gemini users.

      Google has released Lyria 3.5 within the Gemini app and API, allowing the generation of complete songs at 44.1kHz from text or images, complete with SynthID watermarking. On July 31, a Munich court ruled against Suno regarding memorized works and infringing output, claiming jurisdiction over training conducted outside the EU.

      According to Google, Lyria 3.5 is now accessible to all Gemini users worldwide via web and mobile. This model is capable of producing entire songs, including verses, choruses, and bridges. It was first introduced in Google Flow Music on July 29 and is also operational in Google AI Studio and Google Vids.

      The API generates 44.1kHz stereo audio from text or images, according to the documentation. Developers can provide up to ten images, create their own lyrics using section tags, and utilize timestamps to indicate when instruments should enter.

      There are two models available. Lyria 3 Clip consistently creates 30-second clips, while Lyria 3.5 generates tracks that last a couple of minutes. The generation process is single-turn, meaning a clip cannot be edited across different prompts.

      Specific guardrails are in place; safety filters prevent requests for certain artist voices and the creation of copyrighted lyrics. These restrictions closely resemble details from a European court case, although Google has not indicated any connection.

      On July 31, a Munich court determined that Suno had violated copyright in a groundbreaking ruling for Europe. The court held that the works were memorized and reproducibly included in model versions stored on German servers, which excluded them from the text and data mining exception. It also ruled that Suno was accountable for the outputs, rejecting the defense that complex prompts shifted responsibility to the users.

      The court also asserted jurisdiction over training performed outside the EU, conducted its own analysis of US fair use, and concluded that it did not apply. The injunctions prohibit training on the works, retaining them in model parameters, and offering the trained models in Germany. The order included language regarding copying within the United States. This judgment is not final and is subject to appeal.

      The market that Lyria is entering is already crowded, with more than half of new uploads to Deezer being AI-generated. Deezer has been responding to this trend, while Spotify has been less proactive. Approximately 90,000 fully AI-generated tracks are uploaded daily, representing between 1% and 3% of actual listening, with up to 85% of streams in 2025 having been found to be fraudulent.

      Each Lyria track includes a SynthID watermark that Google claims is imperceptible. Suno pursued a similar approach, initiating watermarking following the onset of lawsuits.