Google has launched Lyria 3.5, its latest artificial intelligence model for generating complete songs from text prompts, with upgrades aimed at more natural vocals, stronger lyrics, and greater control over a track’s structure.
The company announced the model on July 29 and began rolling it out through Flow Music. The web platform lets users create, publish, remix, and share songs, and says it is free to start without a credit card.
Lyria 3.5 is designed to generate cohesive tracks of up to three minutes. Google says users can now guide tempo and duration more directly while receiving music with richer melodic structures, better pronunciation, and vocals intended to sound more expressive and emotionally nuanced.
The launch extends Google’s push into generative AI for creative production. Instead of limiting users to short audio samples, Flow Music presents the model as a studio-style system for developing complete songs with lyrics, instrumentation, vocal direction, and production preferences.
Google Improves Vocals, Lyrics, and Prompt Control
The most visible change is vocal quality. Google says Lyria 3.5 produces more realistic singing, clearer pronunciation, and greater emotional variation, addressing areas where AI-generated songs can still sound flat, synthetic, or disconnected from the lyrics.
Lyrics have also been upgraded. The company says the model follows both simple and complex prompts more accurately and has stronger awareness of song structure, helping verses, choruses, and other sections fit together more naturally.
Users can describe a genre, tempo, instrumentation, vocal profile, mood, and lyrical theme. Google’s official prompt guide recommends specifying details such as vocal range, dynamics, instruments, and how the energy should change between sections.
Creative control is becoming a major competitive feature in AI music. Users increasingly expect to direct individual characteristics of a song rather than accept a single finished output generated from a short description.
That shift is also visible in competing releases. ElevenLabs recently introduced Music v2 with section-based composition, improved multilingual vocals, and tools for regenerating selected parts of a track without replacing the entire song.
Flow Music approaches the workflow conversationally. Its Producer interface is designed to let users discuss changes as they would in a studio, while the wider service includes playlists, song publishing, audio effects, stem separation, and AI-generated music videos using Google’s Veo technology.
Google has not published pricing tiers, generation limits, or a detailed regional availability list for Lyria 3.5. The company’s launch announcement directs users to Flow Music, while the platform currently advertises daily credits and free entry without explaining all paid usage conditions on its public homepage.
Lyria 3.5 Uses Latent Diffusion for Audio Generation
A model card published by Google DeepMind describes Lyria 3.5 as a text-to-audio system that produces music and lyrics. The model uses a diffusion model architecture applied to compressed temporal audio representations.
Diffusion systems learn to transform noise into structured output. In music generation, that process can be used to create waveforms and musical patterns that remain coherent across rhythm, melody, vocals, and arrangement.
Google says Lyria 3.5 was trained on audio datasets annotated with text descriptions at different levels of detail. The company used its TPU infrastructure, JAX software framework, and ML Pathways system for training.
The model was evaluated with automated tests and human judgments developed with music specialists. Google compared it with several other music-generation systems, including Lyria 2, and reported significant gains in audio fidelity and improved adherence to prompts involving lyrics.
However, the company did not publish numerical benchmark scores or detailed comparisons with major commercial rivals. That makes it difficult to measure how large the improvement is until creators test the model across different genres, languages, vocal styles, and production requirements.
Google says safety measures were applied during training and deployment, including dataset filtering, supervised fine-tuning, reinforcement learning, red-team testing, and product-level content filters. Every generated track receives an imperceptible SynthID watermark intended to identify audio created or edited with Google’s AI.
The model card does not provide a detailed public list of the recordings used for training. That leaves questions about training-data transparency at a time when musicians, labels, and publishers are closely examining how AI companies obtain and use copyrighted music.
Lyria 3.5 nevertheless strengthens Google’s position in an increasingly competitive AI music market. Suno, Udio, ElevenLabs, and other developers are racing to improve vocal realism, song length, editing controls, and commercial usability.
For users, the immediate difference is practical: Lyria can now generate longer, more structured songs while offering tighter control over how they sound. The larger test will be whether those improvements make Flow Music useful beyond experimentation and into repeatable creative workflows for musicians, video creators, brands, and producers.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
