Google brings its Lyria 3.5 music model to every Gemini user

Google has broadened access to its advanced music generation AI, Lyria 3.5, embedding it into the Gemini app and associated API for users worldwide. This latest iteration allows for the creation of complete songs, featuring structured elements such as verses, choruses, and bridges, all produced at CD-quality 44.1kHz stereo audio.

The rollout follows the initial debut of Lyria 3.5 within Google Flow Music on July 29 and extends its availability to platforms including Google AI Studio and Google Vids. The model supports input from both textual prompts and images, enabling developers to customize lyrics and instrument timing with precision. Two variants of the model serve different use cases: Lyria 3 Clip delivers concise 30-second segments, while Lyria 3.5 can generate multi-minute compositions. However, users should note that generated clips are single-turn outputs without iterative editing across prompts.

A notable feature of Lyria 3.5 tracks is the inclusion of SynthID watermarking technology, which embeds an imperceptible identifier within the audio files. This watermark aims to trace and verify AI-generated content, addressing growing concerns about authenticity and intellectual property rights in the AI music space.

This expansion arrives amidst heightened scrutiny over AI-generated music and copyright infringement in Europe. On July 31, a Munich court issued a landmark decision against the AI music startup Suno, a ruling that underscores the legal complexities surrounding AI training and output. The court found that Suno’s models had memorized copyrighted works and reproduced them in infringing outputs stored on servers within Germany, thus disqualifying them from exemptions typically afforded to text and data mining.

Crucially, the Munich judgment asserted jurisdiction over AI training activities conducted outside the European Union and rejected the defense based on U.S. fair use principles. It held Suno responsible for the unauthorized content produced by its models, dismissing arguments that responsibility shifted to end users based on the complexity of prompting. The injunctions imposed include prohibitions on training with copyrighted materials, retaining such content within model parameters, and offering the resulting models in Germany.

While this ruling is not yet final and may be appealed, it signals an emerging legal framework in Europe that could influence how AI music models are developed, deployed, and regulated. Google’s Lyria 3.5 incorporates safety measures aligned with these concerns, including filters that block requests to imitate specific artist voices or generate copyrighted lyrics, though Google has not publicly linked these restrictions to the Suno case.

The market into which Lyria 3.5 enters is already crowded with AI-generated music. Platforms like Deezer report that over half of new uploads are AI-produced, with approximately 90,000 such tracks arriving daily. Despite their volume, these AI tracks currently represent a modest share of total listening, estimated between 1% and 3%, and a significant portion of their streaming activity has been flagged as fraudulent.

In response to the surge in AI-generated content, watermarking technologies like SynthID have become critical tools for platforms and rights holders aiming to identify and manage the proliferation of AI music. Suno’s recent move to adopt similar watermarking strategies reflects industry-wide efforts to respond to legal challenges and maintain control over AI-generated outputs.

While the full impact of these developments is still unfolding, Google’s global deployment of Lyria 3.5 marks a significant step in making sophisticated AI music generation accessible to developers and creators. At the same time, evolving court decisions in Europe highlight the ongoing tensions between innovation and copyright protection in an era defined by artificial intelligence.