# Lyria 3.5 brings full songs to Gemini

> Google's new music model can turn text or images into structured stereo tracks lasting a couple of minutes, but its quality gains and training-data claims still come from Google.

_Source: Google announcement and developer documentation, independently cross-checked by The Decoder; quality improvements remain vendor-reported · 2026-09-06 · 6 min read · Verified against primary sources_

Canonical: https://iyu.app/e/google-lyria-3-5-gemini-music

## The 60-second version

Lyria 3.5 turns Gemini into a prompt-based music workspace for structured stereo songs lasting up to about three minutes.

**Key points**

- Gemini users can choose genre, vocals or instrumental output, templates and track length.
- The API accepts text or image prompts and returns 44.1 kHz stereo MP3 audio with song structure.
- Google's sound-quality improvements are self-reported and have not been independently benchmarked in the launch materials.
- DeepMind says every generated track carries an imperceptible SynthID watermark.

**Verdict.** This is a meaningful jump in access and song structure, but it remains a creative draft tool that needs human listening, rights checks and disclosure judgment.

## Full explainer

> **⚑ Caveat:** Google describes Lyria 3.5 as its best-sounding model, with more expressive vocals and richer arrangements. Those quality comparisons are vendor-reported; the launch materials do not include an independent listening test or public comparative benchmark.


### What changed — Gemini now has a song generator

Google released **Lyria 3.5** in the Gemini app and Gemini API. In Gemini, a user can describe or select a genre, choose vocals or an instrumental, start from a template and request a shorter or longer track. Google says the rollout is global on the web and mobile app.

The same model is available to developers through Google AI Studio and the Gemini API, and Google also lists Flow Music and Google Vids as places where it can be used. The important product shift is access: music generation is moving from a specialist demo into familiar creation tools.

- **44.1 kHz** — stereo MP3 output specified in Google's API documentation
- **Up to 3 min** — maximum track length stated on DeepMind's model page
- **2 inputs** — music can be prompted with text or an image


### Under the hood — Longer tracks need musical structure

Google's developer guide distinguishes Lyria 3.5 from its fixed 30-second clip model. Lyria 3.5 is intended to generate a song lasting a couple of minutes, with sections such as verses, choruses and bridges. A prompt can influence duration and use timestamps to outline the structure.

- **Prompt control:** Describe genre, tempo, mood, instruments, vocals and song sections.
- **Output:** 44.1 kHz stereo MP3, with generated lyrics and structure available through the API.
- **Typical use:** Backing tracks, demos, jingles, ringtones and early composition sketches.
- **Still required:** Human review for musical fit, awkward lyrics, rights and publication rules.


### Evidence — Better sound is still Google's claim

Google says Lyria 3.5 improves vocals, musicality and arrangements. The launch pages provide examples, but not a blinded listening study or a public benchmark against competitors. That means the feature set and availability are verifiable, while the size of the quality gain is not yet independently established.

> The release proves broader access and longer structure; it does not by itself prove that every genre or language sounds better.


### Provenance — A watermark helps, but does not settle rights

DeepMind says all Lyria tracks receive an imperceptible **SynthID** watermark, designed to indicate that audio was created or edited with AI. It also says filtering and data labeling are used to reduce harmful content and lyrics.

The Decoder reports Google's position that Lyria was trained on licensed content, while noting that the company has not disclosed the underlying dataset in detail. Even with provenance marking, creators remain responsible for checking platform disclosure, commercial-use terms and any rights attached to their inputs and outputs.

> **i** Use the model as a drafting tool: keep prompts, review lyrics and transitions, compare multiple takes, and disclose AI involvement wherever the destination platform or contract requires it.


## Primary sources

- [Google: Create your best tracks yet with Lyria 3.5 in Gemini](https://blog.google/innovation-and-ai/products/gemini-app/better-tracks-lyria-gemini/)
- [Google AI for Developers: Generate music with Lyria 3.5](https://ai.google.dev/gemini-api/docs/music-generation)
- [Google DeepMind: Lyria 3.5 model overview](https://deepmind.google/models/lyria/)
- [The Decoder: Google brings Lyria 3.5 into Gemini](https://the-decoder.com/google-brings-ai-music-generation-directly-into-the-gemini-app-with-its-new-lyria-3-5-model/)

---
_Published by iyu (https://iyu.app) — the day's AI news, checked against primary sources and rewritten in plain language. Free to quote with attribution and a link to the canonical URL._