You asked for "an upbeat pop song," waited for the generation, and got back three minutes of the same eight-bar loop with no real chorus, no bridge, and vocals that trail off. It technically ran, but it is not a song — and it burned a generation you paid for. That gap between "the model worked" and "I got a finished track" is where most people give up on Lyria 3 Pro.
The problem is almost never the model. It is the workflow. Most how-to articles stop at "type a prompt and hit generate," which is exactly the advice that produces loops. This guide walks the full loop the way people who actually ship songs run it: pick the right access path, write the prompt with Google's official formula, force structure with the timestamp technique, set the instrumental/lyrics/duet flags, run a cheap 30-second test on the base model first, then spend a Pro generation and iterate. The steps competitors skip are steps 4 and 5 — and they are the difference between a loop and a song.
What "using Lyria 3 Pro" actually means
Lyria 3 Pro is Google DeepMind's advanced music model. From a text prompt it writes original songs up to 3 minutes long — intros, verses, choruses, and bridges — including AI-generated vocals and lyrics, and it returns high-fidelity stereo audio at 48 kHz. Google announced it on March 25, 2026, about a month after the base Lyria 3 (30-second clips) launched inside the Gemini app.
Two facts shape the whole workflow below:
- A Pro request returns a full ~3-minute song in one generation. You do not stitch 30-second chunks. That is why structure has to be in the prompt up front — you cannot bolt a chorus on afterward.
- Inputs are text prompts, your own lyrics, PDF, and reference images (up to 10). There is no audio upload and no voice cloning. You describe what you want; you cannot hum a melody at it.
Every output also carries an invisible SynthID watermark and supports C2PA content credentials, so a Lyria track is always identifiable as AI-generated.
Step 1 — Pick your access path
You cannot "use Lyria 3 Pro" in the abstract; you use it through a specific product, and the product decides your cost model and setup. There are several official paths, plus credit studios.
| Access path | Best for | Setup / cost model |
|---|---|---|
| Gemini app | Casual creators already paying for Gemini | Needs a paid Gemini subscription |
| Google Vids | Workspace users scoring videos | Google Workspace / AI Pro & Ultra |
| Google AI Studio / Gemini API | Developers, automation | Pay-per-generation, metered in preview |
| Vertex AI | Teams already on Google Cloud | API + Media Studio, public preview |
| ProducerAI | Trying it globally with a free option | Free + paid tiers |
| Credit studios (e.g. lyria3-pro.com) | Anyone who wants Pro with no Google Cloud setup | Pay-as-you-go credits |
If you just want to make a song without wiring up Google Cloud, a subscription, or an API key, a credit-based studio is the lowest-friction route — you buy credits and generate. That is the path this guide assumes for the hands-on steps, but the prompting technique is identical everywhere. If you want to skip straight to it, the Lyria 3 Pro generator on this site runs the model directly, and the pricing page lists exact credit costs.
Rule of thumb: If you touch it daily and script it, use the API. If you touch it weekly and just want songs, use a credit studio or the Gemini app.
Step 2 — Write the prompt with the official formula
Google publishes a prompt formula for Lyria 3 Pro. Fill every slot and the model has far less room to default to a generic loop:
[Genre & style] + [Mood] + [Instrumentation] + [Tempo & rhythm] + [Vocal style & language] + [Lyrics]A worked example:
Dream-pop with shoegaze textures. Wistful and nostalgic, slightly euphoric.
Layered reverb-drenched electric guitars, analog synth pads, soft live drums.
Around 95 BPM, relaxed backbeat.
Female lead vocal, breathy and intimate, in English.Two things people miss here. First, be specific about instrumentation and tempo — "layered reverb-drenched electric guitars, ~95 BPM" gives the model a target; "nice guitars" does not. Second, name the vocal style and language explicitly. Lyria 3 Pro can sing in 8 languages (English, German, Spanish, French, Hindi, Japanese, Korean, Portuguese), and it will not guess correctly unless you say so.
Step 3 — Set the flags for vocals, lyrics, and duets
Three simple flags control how vocals behave. These are the fastest way to fix the two most common complaints ("I wanted an instrumental" and "it ignored my lyrics"):
- Instrumental: end the prompt with
Instrumental.and the model produces no vocals. - Your exact lyrics: prefix your words with
Lyrics:and the model sings them rather than inventing its own. - Duets: describe multiple singers ("a male and female vocal trading lines in the chorus") to get more than one voice.
Lyrics:
Verse 1: We drove past the old town lights...
Chorus: And we never looked back, never looked back...Step 4 — Force structure with the timestamp technique
This is the step that turns a loop into a song, and it is the one most guides never mention. Instead of hoping the model invents an arrangement, you script the arrangement yourself using timestamp cues from Google's official prompting guide. You lay out the song on a timeline with [mm:ss] markers and the model treats each marker as a guaranteed section change:
[00:00] Soft piano intro, sparse, building tension.
[00:20] Verse 1 begins, add bass and light drums.
[00:55] Chorus, full band, anthemic, layered backing vocals.
[01:30] Verse 2, pull the drums back slightly.
[02:00] Bridge, strip to piano and vocal, then rebuild.
[02:30] Final chorus, biggest energy, extra harmonies.
[03:00] Outro, fade to piano.Because a Pro generation returns one continuous ~3-minute track, this map is your only chance to control where the chorus lands and when the bridge hits. Give the model an arrangement and it stops repeating itself. This is the single highest-leverage move in the whole workflow.
Step 5 — Run a cheap 30-second test first
Before you spend a full Pro generation, sketch on the base Lyria 3 model, which returns up to 30 seconds. Use it to check that your genre, mood, tempo, and vocal direction are landing — the vibe, not the full arrangement. If the 30-second sketch already sounds wrong, a 3-minute version will just be wrong for longer.
This is the prototype-then-produce discipline: prototype on Lyria 3, produce on Lyria 3 Pro. A 30-second clip is cheaper (the base model is billed per 30 seconds, versus one full-song charge for Pro per third-party reports — verify current rates), so you iterate on the idea for pennies before committing to the finished track.
Step 6 — Generate, then iterate deliberately
Run the full Pro generation with your formula prompt plus the timestamp map. When it comes back, do not just re-roll blindly. Change one variable at a time:
- Chorus not big enough? Add instructions at that timestamp ("full band, layered backing vocals").
- Vocals too prominent? Adjust the vocal-style line, or switch to
Instrumental. - Wrong energy? Nudge the tempo or swap the instrumentation, keep everything else fixed.
Reference images can also steer mood — attach up to 10 images to set an aesthetic the text alone struggles to convey (a foggy coastline for a melancholic track, neon signage for synthwave).
How to tell it worked
A finished song, not a loop, should pass these checks:
- Distinct sections. You can point to where the verse ends and the chorus starts — they sound different.
- Dynamic arc. Energy rises into the chorus and pulls back for the bridge, rather than sitting flat for three minutes.
- A resolved ending. The track lands on an outro instead of cutting off mid-phrase.
- Vocals that match your brief — right language, right style, and your lyrics if you supplied them.
If it fails on "distinct sections," your timestamp map was too vague or missing — go back to Step 4.
Common mistakes that produce loops
| Mistake | Fix |
|---|---|
| One-line vibe prompt ("chill lofi song") | Fill all six formula slots (Step 2) |
| No arrangement described | Add [mm:ss] timestamp cues (Step 4) |
| Testing ideas on full Pro generations | Sketch on base Lyria 3 first (Step 5) |
| Expecting to upload a melody or clone a voice | Not supported — describe it in text instead |
| Re-rolling randomly when unhappy | Change one variable at a time (Step 6) |
Ready to run the full workflow end to end? The AI music generator here runs Lyria 3 Pro on pay-as-you-go credits, so you can test a 30-second sketch and then produce the full song without a Google Cloud project or subscription. For a deeper background on the model itself, see the complete Lyria 3 Pro guide and the Lyria 3 overview.
FAQ
How long can a Lyria 3 Pro song be? Up to 3 minutes, returned as one continuous track in a single generation. The base Lyria 3 model tops out at 30 seconds.
Why does my Lyria 3 Pro song sound repetitive? You most likely gave it a mood but no arrangement. Use the timestamp technique (Step 4) to script distinct sections — intro, verse, chorus, bridge, outro — so the model changes the song over time instead of looping one idea.
Can I upload my own audio or clone a voice with Lyria 3 Pro? No. Lyria 3 Pro accepts text prompts, your own written lyrics, PDF, and up to 10 reference images. There is no audio upload and no voice cloning.
How do I make an instrumental with no vocals?
End your prompt with Instrumental. and the model generates the track without any sung vocals.
What language can the vocals be in? Lyria 3 Pro sings in 8 languages: English, German, Spanish, French, Hindi, Japanese, Korean, and Portuguese. Name the language in your prompt's vocal-style slot.
Do I need a paid Google plan to use Lyria 3 Pro? Not necessarily. The Gemini app needs a paid subscription, but credit-based studios and ProducerAI let you use the model without a Google account setup. Credit studios like this one charge pay-as-you-go — see the pricing page.
Sources
- Google — Lyria 3 Pro announcement: https://blog.google/innovation-and-ai/technology/ai/lyria-3-pro/
- Google — Build with Lyria 3 (developer): https://blog.google/innovation-and-ai/technology/developers-tools/lyria-3-developers/
- Google Cloud — Ultimate prompting guide for Lyria 3 Pro: https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-lyria-3-pro
- Google Cloud — Lyria 3 & Lyria 3 Pro on Vertex AI: https://cloud.google.com/blog/products/ai-machine-learning/lyria-3-and-lyria-3-pro-on-vertex-ai
- Google AI for Developers — Generate music with Lyria 3: https://ai.google.dev/gemini-api/docs/music-generation
- DeepMind — Lyria 3 model card: https://deepmind.google/models/model-cards/lyria-3/
- TechCrunch — Google launches Lyria 3 Pro: https://techcrunch.com/2026/03/25/google-launches-lyria-3-pro-music-generation-model/
