You typed a prompt, waited, and got back something wrong — a three-minute loop instead of a song, a refused prompt, vocals in a language you did not ask for, or nothing at all. It is easy to conclude that Lyria 3 Pro is broken. It almost never is. Most "Lyria 3 not working" reports trace back to a handful of documented limits and workflow mistakes, and each one has a clean fix.
This is a troubleshooting handbook, not a feature tour. Most articles that rank for this problem either restate the marketing page or tell you to "try again," which does not help when a generation keeps failing the same way. Below, every issue is broken into Symptom / Likely cause / Fix, grounded in what Google actually documents — the 3-minute cap, the lack of audio input, the content checks, the supported languages, and preview access. No invented error codes, no guessing.
Before you troubleshoot: three facts that explain most failures
Three documented facts about Lyria 3 Pro cause the majority of confused reports. Keep them in mind while you read:
- A Pro request returns one full song of up to 3 minutes in a single generation. You do not stitch 30-second chunks, so structure has to be in the prompt up front.
- The only inputs are text prompts, your own lyrics, PDF, and reference images (up to 10). There is no audio upload and no voice cloning. If you are trying to feed it a sound file, that is not a bug — the feature does not exist.
- Outputs are checked against existing content, and the model is trained on licensed data. That is why some prompts get refused — working as designed, not glitching.
Problem 1 — You get a repetitive loop instead of a real song
Symptom: The generation runs successfully, but you get three minutes of the same eight-bar phrase — no distinct verse, chorus, or bridge.
Likely cause: Your prompt described a vibe but not an arrangement. Because a Pro request generates the whole song at once, the model has no sections to change between if you never told it the song has any.
Fix: Use the timestamp technique from Google's official prompting guide. Script the arrangement with explicit cues so the model is forced to change sections:
[00:00] Soft piano intro, sparse
[00:20] Verse 1, add brushed drums and bass
[00:55] Chorus, full band, bright and lifted
[01:30] Verse 2
[02:00] Bridge, strip back to piano and vocal
[02:30] Final chorus, biggest energy
[03:00] Outro, fadeCombine that with Google's prompt formula — [Genre & style] + [Mood] + [Instrumentation] + [Tempo & rhythm] + [Vocal style & language] + [Lyrics] — so every slot is filled. A structured prompt is the single biggest fix for the "it made a loop" complaint; see our Lyria 3 Pro prompt guide for the full walkthrough.
Rule of thumb: If you did not put timestamps or named sections in the prompt, expect a loop.
Problem 2 — Your prompt gets refused
Symptom: The model declines to generate, or returns a generic track that ignores what you asked for. This happens most often when you name a real artist or reference a copyrighted song.
Likely cause: Google states that Lyria's training data is licensed and permissible, and that outputs are checked against existing content. Prompts that ask the model to imitate a named artist or reproduce a known song run into those content checks. This is a deliberate guardrail, not an outage.
Fix: Describe the sound, not the name. Instead of "make it sound like [Artist]," describe the genre, era, instrumentation, tempo, and vocal style that define that sound — for example, "a 2010s stadium pop ballad, breathy female lead vocal, piano and swelling strings, slow 70 BPM, emotional and hopeful." You keep the aesthetic and clear the check.
Rule of thumb: If your prompt contains a proper noun that names a musician or a specific track, rewrite it as a description before you blame the model.
Problem 3 — It will not make a song longer than 3 minutes
Symptom: You asked for a 5-minute track and got roughly 3 minutes, or the request behaved oddly.
Likely cause: Lyria 3 Pro generates original songs up to 3 minutes. That is the documented ceiling for a single generation, not a temporary limit.
Fix: Design within the cap. Three minutes is a full radio-length song, so arrange for it: tighter intro, two verses, two choruses, one bridge, a short outro. If you genuinely need more, generate extra sections as separate prompts and assemble them in an editor. If your track came back much shorter than 3 minutes, you are likely on the base model, not Pro (see Problem 8).
Rule of thumb: Plan the song to land inside 3 minutes rather than fighting the cap.
Problem 4 — Uploading audio or cloning a voice does nothing
Symptom: You want to upload a reference song, hum a melody, or clone a singer's voice, and cannot find the option — or it is ignored.
Likely cause: Lyria 3 Pro accepts text, your own written lyrics, PDF, and reference images (up to 10). It does not accept audio input and does not do voice cloning. There is no hidden setting for this.
Fix: Translate the audio idea into text and images. For mood, describe it or attach a reference image that captures the aesthetic. For a specific vocal character, describe it in words — "warm baritone, intimate and close-mic'd" — in the vocal-style slot. If a real voice clone is a hard requirement, Lyria is the wrong tool and no prompt will change that.
Rule of thumb: If your plan starts with "upload a sound file," stop — reframe it as a description.
Before you burn another failed generation, test somewhere that shows exactly what you are charged and lets you iterate quickly. The Lyria 3 Pro music generator on this site runs the model directly with pay-as-you-go credits, so you can re-run a fixed prompt without wiring up Google Cloud or a subscription.
Problem 5 — The vocals sound wrong or come out in the wrong language
Symptom: You asked for sung vocals in a specific language and got garbled words, an accent you did not want, or a different language entirely.
Likely cause: Lyria 3 Pro sings in 8 languages: English, German, Spanish, French, Hindi, Japanese, Korean, and Portuguese. Request a language outside that list and the results are unpredictable. Vague vocal instructions also let the model default to whatever it likes.
Fix: First, confirm your target language is one of the eight supported. If it is not, that is the cause — choose a supported language or generate instrumental. If it is supported, be explicit in the vocal-style slot: state the language, the vocal character, and the delivery ("clear female vocal, sung in Spanish, warm and expressive"). To supply your own words, use the Lyrics: prefix; for no vocals, end the prompt with Instrumental.
Rule of thumb: Unsupported language equals unpredictable vocals — check the list of eight first.
Problem 6 — An overloaded prompt produces mush
Symptom: You wrote a long, detailed prompt with many instruments and mood swings, and the output is a muddy, incoherent mess.
Likely cause: Too many competing instructions in one flat block. The model cannot prioritize ten contradictory ideas at once, so it averages them into noise.
Fix: Move complexity out of the sentence and into structure. Keep the core formula slots clean, then use timestamps (Problem 1) to introduce different instruments or moods over time instead of all at once — one clear idea per section beats ten in one paragraph. If it is still mush, strip back to genre, mood, tempo, and one lead instrument, then add detail one layer at a time.
Rule of thumb: Add richness across the timeline, not inside a single overloaded sentence.
Problem 7 — It is unavailable, metered oddly, or the price changed
Symptom: The model is not where you expected it, a generation is billed differently than last time, or you hit API access limits.
Likely cause: On the Gemini API and Vertex AI, Lyria 3 Pro is in public preview and billed pay-per-generation, metered. Preview means availability and rate cards can change while Google iterates. On Vertex, the model id surfaces as google/lyria-3-pro-preview. There is no single unified published price.
Fix: Confirm which access path you are on — Gemini app, Google Vids, Google AI Studio / Gemini API, Vertex AI, ProducerAI, or a credit studio — and check Google's current rate card for that path rather than assuming last month's numbers. Any per-song dollar figures circulating online are third-party estimates; verify them against the live rate card. If you would rather not track preview pricing, a credit-based studio bills a flat, visible cost per generation; our pricing page lists exact credit amounts.
Rule of thumb: In public preview, always check the current rate card before assuming a price.
Problem 8 — You are on base Lyria 3, not Lyria 3 Pro
Symptom: Songs top out around 30 seconds, or structure feels thin no matter what you prompt.
Likely cause: You are using the base Lyria 3 model, which produces clips up to 30 seconds with only basic structure, rather than Lyria 3 Pro, which produces full songs up to 3 minutes. The base model is great for sketches and social clips but was never meant to deliver a finished song.
Fix: Verify the model. In a product UI, look for the Pro tier explicitly; on the API, confirm you are calling the Pro model, not the 30-second base model. The intended workflow is to prototype on Lyria 3, then produce on Lyria 3 Pro. For the differences in plain terms, see what Lyria 3 is and how it works and our Lyria 3 Pro overview.
Rule of thumb: If everything caps at 30 seconds, you are on the base model, not Pro.
Quick diagnostic table
| What you see | Most likely cause | First fix to try |
|---|---|---|
| Loop, no sections | No structure in the prompt | Add timestamp cues |
| Prompt refused | Named artist / copyrighted reference | Describe the sound, drop the name |
| Caps near 3 min | Documented Pro length limit | Arrange within 3 minutes |
| Caps near 30 sec | You are on base Lyria 3 | Switch to Lyria 3 Pro |
| Audio upload ignored | No audio input / voice cloning | Reframe as text + images |
| Wrong-language vocals | Language outside the 8 supported | Pick a supported language |
| Muddy output | Overloaded single prompt | Spread detail across timestamps |
| Unavailable / odd billing | Public-preview access & metering | Check the current rate card |
When it is genuinely not you
Sometimes it really is the service. If generations error out or the service is unreachable across several tries and different prompts, that points to a transient problem on your access path — wait and retry, or switch paths. But run one clean prompt with the full formula and timestamps first: that single test rules out the eight causes above.
FAQ
Why does Lyria 3 Pro keep giving me a loop instead of a song?
Because your prompt described a mood but not an arrangement. A Pro request generates the whole song at once, so it needs explicit section cues. Add timestamps like [00:00] to [03:00] to force verse, chorus, and bridge changes.
Why was my prompt refused? Most refusals come from naming a real artist or referencing a copyrighted song. Google checks outputs against existing content and trains on licensed data, so imitation prompts hit a guardrail. Describe the genre, era, and vocal style instead of the name.
Why are my vocals in the wrong language? Lyria 3 Pro sings in eight languages: English, German, Spanish, French, Hindi, Japanese, Korean, and Portuguese. Anything outside that list is unpredictable. Confirm your language is supported and state it explicitly in the prompt.
Why is my song only 30 seconds long? You are almost certainly on the base Lyria 3 model, which caps at 30 seconds. Switch to Lyria 3 Pro for full songs up to 3 minutes.
Fix it in one clean run
Almost every "Lyria 3 Pro not working" case is one of the eight issues above — a loop from a structureless prompt, a refusal from a named artist, the length caps, an attempt at audio input, an unsupported language, an overloaded prompt, or a preview-access quirk. Work down the diagnostic table, rebuild one prompt with the full formula and timestamps, and most problems disappear.
The fastest way to confirm the fix is to re-run that clean prompt somewhere with visible costs and quick iteration. Try it on the Lyria 3 Pro generator — one structured prompt is usually all it takes to turn a failed generation into a finished song.
