Google has launched Lyria 3.5, the latest version of its artificial intelligence music-generation model, bringing advancements across musicality, lyrics, and vocal quality. According to Google DeepMind, the tech giant’s AI research division, the upgraded model delivers tracks that sound more natural alongside vocals featuring more expression and emotion.
Google Launches Lyria 3.5 Inside Flow Music Platform
The model rolled out on a Wednesday inside Gizmodo, the company’s AI song-creation platform, which is available as part of the Google Labs suite. Google previously acquired a music generator startup called Producer in February and rebranded it as Google Flow Music in April.
Key Enhancements in Vocals, Lyrics, and Track Structure
Google reported that Lyria 3.5 brings several technical improvements over prior iterations:
* Vocals: More realistic and emotionally nuanced vocals paired with improved pronunciation. * Musicality: Richer and more complex melodic structures that sound natural. * Lyrics: Higher-quality lyrics featuring better prompt adherence and structural awareness. * Customization: Expanded creative options allowing users to modify specific elements such as vocals, drums, and bass, with track lengths ranging from 30 seconds to three minutes. * Control: Easier control over track tempo and duration.
The system allows users to generate songs based on detailed prompts describing themes or vibes, such as requests for a goofy R&B jam about a lonely sock.
Copyright Silence and Ongoing Legal Scrutiny
Unlike previous releases, the announcement for Lyria 3.5 omitted several elements standard to earlier Google product rollouts. The blog post did not contain a named author, made no mention of how the model was trained, and did not reference SynthID, the watermarking technology Google embeds in AI-generated tracks. It also lacked the copyright framing that accompanied earlier launches.
When Google launched Lyria 3 in February inside the Gemini app to generate 30-second tracks, the company described it as designed for original expression rather than mimicking existing artists, noting it had been mindful of copyright and partner agreements. Google has never specified what music it used to train the Lyria model line, maintaining that its training utilizes material that Google and YouTube have the right to use under their terms of service, partner agreements, and applicable law.
That position is currently being tested in court. In March, a group of independent musicians and songwriters filed an 118-page lawsuit accusing Google of training Lyria 3 on copyrighted recordings pulled from YouTube without permission or payment. Google responded that the lawsuit rests on an unsupported hypothesis that the company trained on the plaintiffs’ specific works.
Evolution of the Lyria Model Line
Google DeepMind first unveiled Lyria in November 2023 in partnership with YouTube, launching it alongside Dream Track, a YouTube Shorts experiment that generated tracks using AI voice clones of artists including Charlie Puth, T-Pain, and Alec Benjamin. A subsequent model, Lyria 2, powered YouTube’s “Speech to Song” tool for turning dialogue into music.
The introduction of Lyria 3.5 and its updates to Musicbusinessworldwide move the platform closer to competing services such as Suno and Udio. Within the broader AI music market, Universal Music Group and Warner Music Group settled copyright lawsuits against Udio and agreed to build licensed AI music platforms, while Sony Music filed a fresh suit on July 20 naming more than 30,000 recordings. Warner has also settled with Suno, whereas Universal and Sony’s legal cases against Suno remain ongoing.
Worth a look
