Google releases Lyria 3.5 with major gains in vocal quality and musicality as copyright lawsuits over its training data mount

Google Ships Lyria 3.5 as the Legal Fight Over Its Training Data Escalates

On July 29, Google released Lyria 3.5, the newest version of its AI music generation model, available in Google Flow Music. The company says the upgrade delivers significant improvements across musicality, lyrics, and vocal quality, including richer melodic structures, lyrics with better structural awareness, and more emotionally nuanced vocals with improved pronunciation 1. The release comes just over four months after Google launched Lyria 3 Pro, which introduced three-minute tracks with verse and chorus customization 2.

The same day, The Verge published a detailed account of artists suing AI companies over training data, with the legal battle against Google's Lyria engine front and center 3. Google is upgrading Lyria's output quality at the exact moment musicians are trying to restrict the data pipeline that built it.

Sam Kogon is the lead plaintiff in the ongoing suit against Google's Lyria AI music engine. He and other independent musicians accuse Google of violating its own terms of service by using Content ID data and YouTube content to train Lyria and ProducerAI 3. Kogon told The Verge that Google is "devaluing our work" by "giving it away to people, for now, for free" 3.

Google has filed a motion to dismiss, arguing that YouTube's terms of service grant the company broad rights to "reproduce, distribute, [and] prepare derivative works" from user uploads 3. Google spokesperson Jack Malon told The Verge that the company uses content uploaded to YouTube to improve the product experience across YouTube and Google 3.

The case has an unusual legal posture. Many suits against AI companies allege direct copyright infringement, but the Lyria plaintiffs argue that Google violated its own terms of service. Their contention: uploading a song to YouTube never constituted consent to train AI models, and Google retroactively repurposed user content for a technology that did not exist when the uploads happened 3. Entertainment and IP lawyer Krystle Delgado told The Verge that YouTube's uploader agreement grants an "irrevocable perpetual license," meaning creators cannot revoke Google's access to their content once it is uploaded 3.

Google's blog post announcing Lyria 3.5 makes no mention of the litigation 1. The company can ship a model with more natural vocals and tighter lyrical coherence today. It cannot answer the question of whether it had the legal right to build that model on other people's recordings. Every product improvement deepens the contradiction: the better Lyria gets, the more artists stand to lose, and the more Google depends on training data whose legal status is unresolved.

For builders working in AI music and adjacent creative domains, the Lyria case will shape how they acquire training data. If Google's terms-of-service argument succeeds, platforms hosting user-generated content gain broad training rights from their own ecosystems. If the musicians win, AI companies face a reckoning for models already deployed in their products. Either ruling rewrites the economics of AI-generated content.

Illustrators, authors, and musicians have filed suits against Meta, Anthropic, Stability AI, Midjourney, and Suno, an AI music generator, over AI training practices 3. Several of these cases have been working through the courts since early 2023, and artists told The Verge they are optimistic about how their individual cases will play out 3.

For investors, models trained on contested data carry a liability that balance sheets do not yet account for. For artists like Kogon, the stakes are immediate. He describes making AI-generated music as "the most anti-human thing you could do" 3.

Google's position in court is that YouTube's terms of service already granted those training rights. The courts will decide whether that holds. In the meantime, the model keeps getting better.