Google shipped Lyria 3.5 to Flow Music on July 29th, putting DeepMind's newest music model inside the AI music studio built by Seth Forsgren and Hayk Martiros before Google acquired their startup earlier this year.
https://x.com/googleflowmusic/status/2082496617664413721?s=46 In a thread on X, Google said Lyria 3.5 produces more expressive singing, richer arrangements and tracks that follow creative instructions more closely. The release also adds direct BPM control, giving musicians a way to specify tempo instead of repeatedly generating songs until the model lands near the intended pace.
The Lyria product page identifies 3.5 as the latest model in the family and says it supports variable song lengths of up to three minutes. Google's demonstrations span several languages and genres, with prompts specifying tempo, key, vocal placement, instrumentation and production style.
DeepMind's model card covers Lyria 3 and subsequent versions. It describes a text-to-audio system that applies latent diffusion to temporal audio representations, with text prompts as inputs and music and lyrics as outputs. The model card says Google evaluates the family for audio fidelity, musical quality, vocals and prompt adherence.
From a hobby project to Google's music product
Forsgren and Martiros met as Princeton undergraduates and played together in an amateur band for about a decade. In December 2022, they released Riffusion, a hobby project that used a version of Stable Diffusion to generate spectrogram images and convert them into short audio clips.
The project drew enough usage to become a startup. Riffusion raised $4 million in October 2023 from Greycroft, South Park Commons and Sky9 Capital, then evolved into ProducerAI, a conversational system for generating and revising complete songs.
Google acquired ProducerAI on February 24th and moved its personnel into Google Labs and Google DeepMind. Forsgren now serves as group product manager for Google Flow Music. Martiros, who spent nearly a decade working on autonomous drones at Skydio before building Riffusion, is listed among the contributors to the Lyria model family.
Google initially described ProducerAI as a product powered by Lyria 3, Gemini, Veo and its Nano Banana image model. It later renamed the service Flow Music, connecting the founders' conversational editing interface directly to Google's model portfolio.
Control is the product
Lyria 3.5's BPM setting fits the strategy Forsgren and Martiros have pursued since Riffusion: turning generation into an iterative production process. Flow Music already lets users revise song structure, alter a mix, replace selected sections, create covers and extend tracks through natural-language instructions. Google's product documentation gives examples such as shortening an introduction, adding a bridge, changing instruments or making a kick drum more prominent.
That workflow gives Google a route beyond one-shot text prompts. A musician can generate a draft, select a section and ask the Producer agent to replace it while preserving the rest of the track. Flow Music also supports uploaded audio and images, downloadable stems, music videos and user-built Spaces that function as instruments, effects or sequencers.
Google is already charging directly for the product. Flow Music's current pricing includes a free tier and paid plans ranging from $6 to $48 per month. The paid plans increase generation credits and concurrency, while the highest tier includes early access to new features.
Distribution is moving through industry partnerships as well. On May 6th, Google announced an agreement with Believe and TuneCore to give selected artists, producers and songwriters access to Flow Music. The companies also created a weekly feedback program intended to bring working musicians into product development.
The copyright fight follows every model release
Lyria 3.5 arrives while Google's music-generation work faces a proposed copyright class action. A complaint filed on March 6th alleges that Google used 44 million clips representing 280,000 hours of the plaintiffs' music to train Lyria without permission, according to Bloomberg Law. Those figures are allegations from the plaintiffs and have not been established by a court.
Google has said its music models use material that Google and YouTube have the right to use under their terms of service, partner agreements and applicable law. The company also embeds generated tracks with SynthID, its watermark for identifying AI-generated or AI-edited content.
The model upgrade advances the product Forsgren and Martiros began as an open-source experiment less than four years ago. Under Google, that project has become a paid creation suite, a distribution channel for DeepMind models and a test of whether conversational editing can make generated music useful beyond the first prompt.