Google Flow Music Gets Lyria 3.5, and Now the AI Writes the Words Too
Google DeepMind made it official on July 29: Lyria 3.5 is here, a new AI music generation model rolling out inside Google Flow Music.
The release pushes forward on several fronts at once — musicality, lyrics, vocals, and creative control.
Also Read: Can €10 Billion Close Europe’s AI Compute Gap? Brussels Just Placed the Bet
It’s the most substantial update to DeepMind’s generative music stack since Lyria first debuted. And it drops the company straight into competition with a fast-moving field of AI audio startups.
Lyria 3.5 Arrives In Google Flow With 5 Simultaneous Gains
Lyria 3.5 targets five distinct dimensions of music generation at once. DeepMind announced on July 29 that the model delivers improvements across musicality, lyric quality, vocal performance, creative control, and overall output coherence.
That scope matters.
Most prior generative music systems traded off one quality against another. A model that improved melodic structure often produced weaker vocals; one optimized for lyrics often struggled with timing.
Delivering gains across all five simultaneously is the architectural claim at the center of this release.
Google Flow Music, the platform hosting Lyria 3.5, is DeepMind’s consumer-facing creative suite for AI-assisted audio production. It is designed for both professional producers and casual creators, letting users generate original tracks from text prompts, edit individual stems, and layer AI-generated vocals over custom chord progressions.
What Lyria 3.5 Actually Is
AI music generation works by training a neural model on large corpora of audio and associated metadata.
The model learns statistical relationships between musical elements, instruments, tempos, chord structures, and lyrical forms. On inference, a user provides a prompt or a partial composition, and the model completes or generates audio that matches the learned patterns.
Lyria 3.5 sits in a class of models that go beyond simple loop generation.
It produces full compositions with dynamic structure, meaning a generated track can include an intro, verse, chorus, and bridge rather than repeating a fixed segment. The vocals dimension is particularly technically demanding because singing requires the model to synchronize phonetic timing, pitch contour, and lyric semantics simultaneously.
Creative control refers to the degree to which users can steer generation without the model drifting.
Earlier versions of Lyria, and competitors such as Suno and Udio, have faced criticism that their outputs are stylistically generic or unpredictable. DeepMind’s claim is that Lyria 3.5 inside the platform maintains user intent more reliably across a full track.
A Fast Market DeepMind Can No Longer Afford To Trail
The AI music generation space accelerated sharply through 2025 and into this year.
Suno raised at a valuation above $500 million. Udio attracted major label investment. Meta released its MusicGen family as open weights. Apple integrated Instrument Studio with AI composition tools in Logic Pro.
Google has significant advantages in compute, audio research depth, and an existing creator platform through YouTube.
Lyria 3.5 is the company’s clearest signal yet that it intends to compete for the professional creator market, not just the novelty consumer segment.
The embedding of Lyria 3.5 inside Google Flow rather than releasing it as a standalone API or research artifact also signals a product-first strategy. DeepMind is routing creators through a Google-controlled surface, which keeps usage data inside the company’s ecosystem.
From DeepMind Research Lab To Creator Product
DeepMind’s music AI research dates to its earlier work on WaveNet, the neural audio synthesis model it published in 2016.
That model, originally designed for text-to-speech, demonstrated that deep learning could produce realistic human-sounding audio at the waveform level. It became the foundation for Google Assistant’s voice and set a technical template the broader audio AI field followed.
Lyria emerged from that lineage as a music-specific generation system.
The first public Lyria demonstrations appeared in late 2023, embedded in early versions of the platform and in a collaboration with YouTube for AI music experiments. Those initial releases were notable technically but limited in practical creative utility.
Users could generate short clips with a broad style prompt but had little fine-grained control.
Lyria 3.5 represents the shift from research demonstration to working creative product within Google Flow.
The jump from Lyria to Lyria 3.5, skipping an integer version, mirrors the naming cadence DeepMind has used in other product lines and likely reflects a non-linear improvement rather than an incremental one.
Why The Vocal Upgrade Is The Hardest Part
Of the five areas DeepMind cites, vocals are the most technically demanding and commercially important.
Lyric generation in text is a solved problem for large language models. Melodic structure can be encoded through music theory rules.
But AI singing requires a model to bind all of these together in real time across a waveform.
Current AI vocal models frequently produce what engineers call the “uncanny valley” effect in audio: pitch and rhythm are technically correct, but the phrasing lacks the micro-variations human singers introduce naturally. DeepMind’s claim of improved vocal performance in its creative suite is the hardest to verify without hands-on testing.
It is also the capability that most directly determines whether creators use the tool and Lyria 3.5 for publishable output or only for early-stage ideation.
If the vocal improvement holds up in independent testing, Lyria 3.5 would represent a meaningful threshold crossing for AI-generated music entering commercial production workflows.
Read Next: Anthropic Shuts Claude Out of China: The World’s Largest AI Market, Forfeited
