What is Google Lyria, and can it make a full song with your lyrics?

You open Gemini, pick Create music from the tools menu and type: "a warm acoustic song for my brother's 40th, mention the fishing trip and the canoe that sank". A minute later there is a track with a voice on it, a chorus that comes back, and cover art. It sounds finished. Then you listen a second time. The chorus rhymes "canoe" with "dreams come true", your brother's name never appears, and the second verse is about a sunset nobody mentioned.

That is Google Lyria in one scene. The model sings, it arranges, and inside Gemini anyone with a Google account can try it. Whether it makes your song comes down to a question most people skip: who writes the words.

The short answer to the title is yes, with conditions. The current model, Lyria 3.5, will sing lyrics you supply if you paste them in with section tags, and in the Gemini app a full track runs about two to three minutes. The conditions are a language list, a length cap, a usage limit Google does not publish as a number, and the fact that the description and the words share one text box.

Lyria is a model family, not an app

Nobody downloads "Lyria". It is a family of music generation models from Google DeepMind, and you reach it through several doors Google built around it. DeepMind first presented Lyria in November 2023 alongside music experiments on YouTube. Since then the family has grown to Lyria 2, Lyria 3, Lyria 3.5 and Lyria RealTime, plus an open model called Magenta RealTime.

The confusion starts because each door runs a different model with different limits. Here is the map at the time of writing (October 2026), every row taken from Google's own pages:

Two things in that table matter more than the rest. Only one row is a consumer product with a button: the Gemini app. Everything else is a developer surface or a separate creative tool. And the RealTime model does not sing words at all. Its documentation says it generates instrumental music only; one mode adds vocalizations as another instrument, which is not the same as singing a lyric. If a RealTime demo is what sold you on Lyria and you want a song with your lyrics, you are looking at the wrong model.

If you have not picked a music generator yet and are comparing the whole field, our breakdown of which AI music generator fits which job is the better page for you. This one is for the reader who already has Gemini open and wants to know what Lyria will do with a specific set of words.

Making a song in Gemini, click by click

Google's help page describes the flow, and it is short:

1. In the Gemini text box, open Upload & tools and choose Create music.
2. Optionally upload images or video as extra context for the track.
3. Pick a genre from the list or describe one, and choose vocals or instrumental.
4. Choose the length: short (about one minute) or full (about two to three minutes). The help page says a full track needs the model picker switched to Pro.
5. Write the prompt. Templates exist for jobs like background music and birthday songs.
6. Download the result as an MP4 with the cover art or as an audio-only MP3, or copy a share link.

Every track comes with cover art from Google's image model and carries a SynthID watermark, an inaudible marker that identifies it as AI-generated. Google also says there is a limit on how many tracks you can create: you get a notice when you hit the ceiling, and a choice between upgrading your plan and waiting for the limit to refresh. The help page gives no number, so we will not invent one. If volume matters to you, find your own ceiling on day one instead of discovering it on deadline day.

You need to be 18 or older and signed in with a personal Google account or a supported work or school account. Google's Gemini overview says music generation is offered in every country where the Gemini app is available.

Yes, it can sing words you wrote

Google documents this in two places. The Gemini overview page tells you to describe a topic, add personal details, or "provide your own text with structure tags". The Gemini API documentation says you can write your own lyrics into the prompt, and that tags like [Verse], [Chorus] and [Bridge] help the model follow the song's structure.

This is where the birthday scene gets fixed. When you type only a description, Lyria writes the lyric itself, and a model writing for a stranger reaches for the safest rhyme in the drawer. When you paste a finished lyric, the model has something specific to sing.

A prompt that gives it a real chance looks like this:

Warm acoustic folk, male baritone, fingerpicked guitar, brushed snare, slow build.
Sing these lyrics exactly as written.

[Verse 1]
Forty summers and you still check every knot twice
You still call the lake by the name Dad gave it
The canoe went under back in two thousand nine
You laughed so loud the whole shore heard it

[Chorus]
Mike, the water never wins
It only borrows you a while
Mike, the water never wins
You come home soaking, with that smile

[Verse 2]
Forty summers and the boat is fibreglass now
Your kids row it out to the very same spot
They ask about the old one, you tell them how
It sank with half the bait and all the luck we got

[Chorus]
Mike, the water never wins
It only borrows you a while
Mike, the water never wins
You come home soaking, with that smile

Three things in that block carry the load. The style line sits on top and stays short, one breath of description. The instruction "sing these lyrics exactly as written" draws a line between what to perform and what to sing. And the tags mark the sections, so the chorus returns as a chorus instead of drifting into a third verse. The name opens the chorus and comes back on its third line, where it cannot get lost. Two verses and a chorus sung twice is also a shape that has a fair chance of fitting a two-to-three minute window, which matters in the next section.

Note what Gemini does not give you: separate fields. The description and the words share one box. Suno, a different company, splits them into a Style field and a Lyrics field in its custom mode, so a stray adjective in the description cannot end up sung. In Gemini that separation is your job, and the "sing exactly" line is how you do it. Google's Vertex documentation also lists negative prompting as unsupported for Lyria 3, so "no drums" has no dedicated place to go. Describe what you want to hear instead: "acoustic guitar and voice only".

Where Lyria stops

Lyria 3.5 in Gemini makes a convincing first draft of a song. Four limits decide whether a draft is all you need.

Length. A full track is about two to three minutes. The Lyria 3 Pro preview on Vertex caps at 184 seconds. A lyric with three verses, a bridge and a chorus sung four times will not fit, and if you do not trim it, the model decides what to drop. Cut it yourself so the cut lands where you want it.

Language. Vertex lists the languages for Lyria 3 as English, German, Spanish, French, Hindi, Japanese, Korean and Portuguese. The Gemini API page for Lyria 3.5 says lyrics come out in the language of your prompt and the vocal style adapts to it, but it publishes no list. If your song is in a language outside those eight, run one short track before you plan anything around it.

Volume. Limits apply and no number is published. For one birthday song that never matters. For a channel that needs forty tracks a week, it is the first thing to test.

Control after the fact. In the Gemini app you write a prompt and get a track. If the bridge is perfect and the second chorus is wrong, Google's help page describes no way to fix one section and keep the rest, so the usual move is a new generation. Tools built only for songs tend to give you more handles here, which is the honest reason some people use Gemini to sketch and another tool to finish.

On rights: Google's Terms of Service say that when its services let you generate original content, "Google won't claim ownership over that content". The same terms forbid misleading people into thinking generative AI content was made by a human. Whether anyone owns an AI track, and who, depends on your country, which is a separate question covered in who owns AI-generated music. This is not legal advice.

Lyria or a dedicated song generator

There is no winner here, only a fit.

Gemini with Lyria fits you if you already use Gemini, you need one track this week, your song is in one of the documented languages, two to three minutes is enough, and you are happy with one box for both style and words. Uploading a photo for context and the ready cover art save steps if the track is going straight to a phone or a group chat.

A dedicated song generator fits you if the song is long, if you want the lyric in its own field so nothing from the description leaks into the vocal, or if you expect to iterate on one section. Suno is the one our tools are built around. Our page on Suno alternatives covers the rest of that shelf and who each one suits.

The API fits you if you are building a product. The lyria-3.5 model returns MP3 by default or WAV on request, at 44.1 kHz stereo, and the clip model always returns 30 seconds, which is useful for previews and loops.

The step that stays the same in every one of those paths is the lyric. Every tool on that shelf, Lyria included, sings better words when you hand it better words.

Where sunomarket fits

sunomarket is an independent site. We are not Google and not Suno, and we have no partnership with either. We do not generate audio, in Lyria or anywhere else.

What we make is the part the opening scene got wrong. The lyric generator takes a short form brief (theme, point of view, emotion, structure, hook type, vocal, language) and returns a title and a full lyric with [Verse] and [Chorus] tags already in place. You write "brother's 40th, the canoe that sank in 2009, he laughs at everything, name Mike" into the theme field and get back a lyric that is about Mike, not about dreams coming true.

It is built for Suno's format, and that is where it has been tested. A tagged lyric is plain text, so you can paste it into Gemini's box too, and Google's own docs say section tags help Lyria follow the structure. Tag handling differs from one tool to another, though, so treat the first Lyria track as a test, not a delivery.

Pricing, in dollars: one lyric on the SM1 model costs 20 credits, which is twenty cents at 100 credits to the dollar. SM2 and SM3 cost 80 credits and run an extra pass that flags weak lines and rewrites them. There is no free trial of the generator: a new account starts at zero credits. A plan is $5 a month for 500 credits, or $19 once for a lifetime 1,500 credits that never expire. Details are on the pricing page.

The workflow, start to finish

1. Write the brief before you open any tool. Who the song is for, one concrete detail only you know, the name, the mood. A brief with a sunken canoe in it beats a brief with "happy and nostalgic".
2. Get the lyric. Write it yourself or generate it. Either way, read it out loud once: if a line trips your tongue, it will trip the singer.
3. Trim it to fit. For a full Gemini track, two verses and a chorus that returns two or three times is a safe shape. Put the name or the key phrase in the chorus.
4. Build the prompt. One short style line, then "Sing these lyrics exactly as written", then the tagged lyric. Keep descriptions out of the lyric block.
5. Generate one short track first if your language is not on Google's list or if you have never used Lyria before. It costs you one slot of your limit and saves you a wasted full one.
6. Listen for three things only. Is the name sung correctly? Does the chorus come back as a chorus? Did the model invent any line you did not write? If one line gets mangled, rewrite that line and regenerate. Do not rewrite the whole prompt for one bad word.

If the result is close but you want more control than one text box gives, take the same lyric to a dedicated song generator. The words do not change between tools. Spend your time there first, and whichever generator you open, start from a finished, tagged lyric rather than a description and a hope.