Google released Lyria 3.5 in Flow Music on July 29, 2026, upgrading the model’s vocals, lyrics, and tempo controls. The announcement carried no benchmark and no watermarking detail.
Quick Summary – TLDR:
- Google shipped Lyria 3.5 inside Google Flow Music on July 29, 2026, calling it its newest music generation model.
- Vocals are the headline change, with Google citing more expression, more emotion, and better pronunciation.
- Lyria 3 Pro, the previous release, lifted the track ceiling to three minutes from 30 seconds.
- Sony Music has removed 135,000 AI-generated songs that were falsely attributed to its own recording artists.
- Google’s launch post did not state pricing, regional availability, or whether SynthID watermarks Lyria 3.5 output.
What Happened?
Google launched Lyria 3.5 in Google Flow Music, the company’s music creation tool, on the day it announced the model. A user writes a prompt and the model returns a track. Google grouped the changes into four areas: musicality, lyrics, vocals, and creative control.
The company described each area in its own words, without measured results. Google said Lyria 3.5 lets users “create richer, more complex melodic structures that sound more natural,” and generate “higher quality lyrics with improved prompt adherence and structural awareness.“
Expecting A Minor update? We prefer to B Major. 🎶@GoogleDeepMind just released Lyria 3.5, and we’ve plugged it directly into @GoogleFlowMusic to power your next session.
— Google Labs (@GoogleLabs) July 29, 2026
What this means for your tracks:
✨ Better prompt adherence: Flow Music understands your creative… pic.twitter.com/mXikwIqC6I
Vocals and Lyrics Carry the Upgrade
Vocals drew the most specific language in the release. Google said the model brings “more realistic and emotionally nuanced vocals, plus improved pronunciation,” and that users can “more easily control the tempo and duration of your outputs.“
The previous model sets the baseline for what changed. Google shipped Lyria 3 Pro on March 25, 2026, lifting the track ceiling to three minutes from 30 seconds and letting users define intro, verse, chorus, and bridge sections. Every Lyria 3 Pro output carried a SynthID watermark.
Google has not said how the new duration control interacts with that three-minute ceiling.
A Realism Gain Lands in a Fraud-Heavy Market
Vocal realism is the axis where synthetic music becomes hard to police. Human detection accuracy for deepfakes can fall to 24.5% on high-quality media, according to AI voice cloning fraud data, which also puts the global AI voice cloning market at $4.06 billion in 2026 on a Research and Markets projection.
Streaming platforms are already absorbing the volume. Sony Music removed 135,000 AI generated songs falsely attributed to its artists, and Spotify opened a beta artist approval tool that lets artists review releases before they appear under their names. AI reaches music platforms from other directions too, including Spotify’s Claude-powered recommendations.
Artists with catalogs on major services can check today whether their distributor or platform offers a release approval control and switch it on. It helps reduce the risk of a synthetic track landing on their profile, and it does nothing about uploads posted under a different name.
What the Launch Post Left Out?
Google’s announcement runs to four bullet points and a rollout line. Five things it does not answer:
- Whether SynthID watermarking applies to Lyria 3.5 output, as Google confirmed for Lyria 3 Pro.
- What Google Flow Music costs, or whether Lyria 3.5 sits behind a paid tier.
- Which countries receive the model, and on what schedule.
- What the new duration control tops out at.
- Whether any listening test compared Lyria 3.5 against the previous model.
Google published no evaluation data alongside the release. “More realistic” and “higher quality” are the company’s characterizations of its own model, and no outside party has measured either one.
SQ Magazine’s Takeaway
Lyria 3.5 moves the two variables that decide whether a generated song passes as human. Google’s own list points at both, promising better pronunciation and more emotional nuance than the model delivered before, alongside direct control over tempo and length. That aims the release at creators who want a finished song out of the tool.
What comes next sits with the platforms more than the model. Detection and attribution tooling on streaming services was built against an earlier generation of output, and a vocal upgrade tests it directly. Two things are worth watching: whether Google states that SynthID covers Lyria 3.5, and whether streaming services say how their screening handles the new output.
