---
title: "Can Google Lyria 3 sing in Romanian and play Romanian genres? What we measured through the Gemini API"
description: "Romanian is not on Google's list of supported vocal languages for Lyria 3, but in our tests it worked: about 85% of the words we sent came back as written, diacritics were pronounced correctly, and the intonation was odd in places. Functional instrumentals such as lofi, sleep and ambient worked every time, while Romanian dance names like sârbă did not: the model needs the rhythm described in musical terms. The list price of $0.04 or $0.08 per song understates the real cost of a usable vocal track, because there is no iterative editing."
date: 2026-10-19
lang: en
canonical: https://goingstellar.ro/research/google-lyria-3-romanian-music-test/
author: Going Stellar
publisher: GOING STELLAR S.R.L.
topic: Measurement / AI music
license: CC BY 4.0
alternate_ro: https://goingstellar.ro/ro/cercetare/google-lyria-3-muzica-romaneasca-test/
---

# Can Google Lyria 3 sing in Romanian and play Romanian genres? What we measured through the Gemini API

*Romanian is not on Google's list of supported vocal languages for Lyria 3, but in our tests it worked: about 85% of the words we sent came back as written, diacritics were pronounced correctly, and the intonation was odd in places. Functional instrumentals such as lofi, sleep and ambient worked every time, while Romanian dance names like sârbă did not: the model needs the rhythm described in musical terms. The list price of $0.04 or $0.08 per song understates the real cost of a usable vocal track, because there is no iterative editing.*

> Measurement · AI music · 2026-10-19 · https://goingstellar.ro/research/google-lyria-3-romanian-music-test/

## Key points

- Romanian is not one of the eight vocal languages Google lists for Lyria 3, but in our tests it worked, with about 85% of lyrics sung as written and correct diacritics.
- Functional instrumentals (lofi, sleep, ambient) were usable eight times out of eight; Romanian dance names were not understood, and sârbă came out as horă twice out of two.
- Describe the rhythm, tempo, accompaniment and structure in musical terms instead of naming the dance; Google's genre list for Lyria contains no Romanian genre.
- Lyria 3 Clip costs $0.04 and Lyria 3 Pro $0.08 per song with no free tier, but a usable vocal song cost about ten times that in our sessions, because there is no iterative editing and changed lyrics are billed as normal.
- Lyria 3 is now a legacy preview family: Lyria 3.5 became generally available on 3 September 2026 at $0.08 per song, and our results may not transfer to it.
- The Gemini API terms require users to be 18 or older and exclude apps aimed at minors; paid use keeps prompts and outputs out of product improvement.

## The short answer

Lyria 3 is Google's music generation model family in the Gemini API. It comes in two versions: `lyria-3-clip-preview` for 30-second clips and `lyria-3-pro-preview` for full songs. Google lists eight vocal languages for Lyria 3, and Romanian is not one of them. In our tests Romanian still worked: the model sang real Romanian words, pronounced the diacritics correctly and kept roughly 85% of our lyrics intact. The voice had odd intonation in places.

Functional instrumentals came out well every time we tried them. Romanian folk dance names did not: asked for a sârbă, the model played a horă both times. Describing the rhythm in musical terms is the workaround. For vocal songs with fixed lyrics, the real cost per usable track is on the order of ten times the list price, because a bad verse means regenerating the whole song.

One change since our tests matters. On 3 September 2026 Google released `lyria-3.5` as generally available, and the documentation now calls Lyria 3 a legacy family. Our measurements were made on Lyria 3 and may not carry over to Lyria 3.5.

## Which models exist today

Google launched Lyria 3 in the Gemini API on 25 March 2026 with two models: `lyria-3-clip-preview` for 30-second clips and `lyria-3-pro-preview` for full-length songs. Both are still listed as preview models, and Google's deprecations page shows no shutdown date for either.

On 3 September 2026 Google announced `lyria-3.5` as generally available, with improved musical coherence, natural vocals and finer control of duration and structure. The models page now describes Lyria 3 Pro as the previous generation, and the deprecations page names `lyria-3.5` as its recommended replacement. The clip model has no newer version: `lyria-3-clip-preview` is still the 30-second option in the current guide.

Pricing on Google's page is per request. Lyria 3 Clip Preview costs **$0.04** per song, Lyria 3 Pro Preview costs **$0.08** per song, and Lyria 3.5 costs **$0.08** per song. There is no free tier for any of them.

## What the documentation promises

**Length.** The clip model always produces 30 seconds. Google's Cloud blog says Lyria 3 Pro supports compositions up to three minutes. The longest full song we generated ran 2 minutes 25 seconds.

**Languages.** Google's April 2026 prompting guide for Lyria 3 lists eight vocal languages: English, German, Spanish, French, Hindi, Japanese, Korean and Portuguese, with more languages coming soon. The current Lyria 3.5 guide no longer gives a list. It says the model writes lyrics in the language of the prompt and adapts its vocal style and pronunciation to that language.

**Watermark.** All generated audio carries a SynthID audio watermark, according to the music generation guide. The Cloud blog adds that outputs support C2PA content credentials.

**Editing.** The guide says plainly that music generation is a single-turn process, and that iterative editing or refining a result through further prompts is not supported. Results can also differ between calls with the same prompt.

**Filters.** Every prompt goes through safety filters, including checks for requests for specific artists' voices and for copyrighted lyrics. Prompts that trigger the filters are blocked.

## Romanian vocals: what we measured

We sent Romanian lyrics to `lyria-3-pro-preview` and compared what was sung with what we wrote. About 85% of the words came back as written. The rest were replaced by words that do not exist or were not in our text. Words with ș, ț, ă, â and î were pronounced correctly when we sent them with diacritics. Our listener rated the voice as clearly synthetic in places, with stresses landing on the wrong syllable.

We also compared two ways of writing the lyrics: normal Romanian with diacritics, and a phonetic spelling based on English reading rules, for example *chass* for *ceas* and *dor-u-le* with hyphens to force syllables. One track in each version was judged good, both with the same remark about odd intonation in places. The A/B had no clear winner, so we now start with plain diacritics and only respell the specific words that come out wrong.

Two habits helped. We kept each lyric block to four lines or fewer, and we put the lyrics after a `Lyrics:` header, which Google's own guide recommends, so the model does not sing the stage directions. Shouted interjections, the strigături typical of Romanian party music, came out natural.

## Romanian dances: describe the rhythm, not the name

Google's genre keyword list for Lyria includes polka, cumbia, bhangra and Celtic folk, but no Romanian genre. Our results fit that gap. Asked for a sârbă, the model delivered a horă twice out of two attempts. A horă asked for by name had an odd rhythm, missing the repeating swaying cycle that makes the dance. A lăutărească sounded right in timbre but lacked the usual arc of the genre, which builds and speeds up section by section. The model seems to map the dance names onto a generic folk feel.

The fix is to describe what a musician would hear. These are the formulations we now use. They come from our listener's description of what was missing; we have not yet measured them in a second batch.

**Sârbă:** *very fast, driving dance around 150 BPM, relentless off-beat oom-pah accompaniment, continuous rapid running melody, no tempo changes*.

**Horă:** *moderate circle dance around 100 BPM, cyclical swaying phrase structure that rocks forward and back, strong repeating rhythmic cycle*.

**Lăutăresc dance piece:** *clear dramatic progression: slow expressive rubato intro, then a moderate groove, then accelerating section by section to a frenzied virtuosic finale*.

We left time signatures out on purpose. Written descriptions of these dances disagree on the meter, and we have not tested which notation the model follows better.

## Genre decides how synthetic the voice sounds

The functional genres were the easy case. All eight lofi, sleep and ambient instrumentals we generated were judged usable. There is no voice to expose, and the genres tolerate repetition.

With vocals, genre mattered more than the language. A ballad came out as a genuine ballad, but it is the hardest test for a synthetic voice: the slow tempo and exposed singing leave every odd stress and missing breath audible. A manea with Romanian lyrics sounded acceptable because heavy auto-tune and reverb are the norm in that genre, and the same processing hides the model's artefacts. We expect the same for trap, where auto-tune is also part of the aesthetic; that is a judgement from our listening, not a measurement across genres.

## What a usable track really costs

The list price is the cost of one generation, not of one usable song. Because there is no iterative editing, a single bad verse means throwing the whole track away and generating it again at full price. In our sessions, vocal songs with fixed Romanian lyrics needed on the order of ten attempts for one we would keep, so the real cost per usable track was roughly ten times the list price. Functional instrumentals were the exception: there, the first take was usually usable.

Lyric changes are the second cost. When a word is sung differently from what we wrote, the API returns a normal successful response and the request is billed like any other. There is no warning. The only check is to listen to the whole track against the text. Our notes describe whole lines being replaced by generic ones when lyrics touched a filter; Google's documentation only says that prompts triggering the filters are blocked, so we could not confirm that mechanism. With Lyria 3.5 the response also includes the generated lyrics as text, which makes a first automatic comparison possible.

Google's own guide suggests a cheaper loop: test prompts on the 30-second clip model, at half the price, before committing to a full song. We do the same for genre and rhythm wording.

## Rate limits, terms and why we use the paid API

Google publishes no Lyria-specific rate limit. Its rate limit page says limits apply per project, can be checked in AI Studio and are tighter for preview models. Our project ran at about 10 requests per minute, so batches need a pause and retry.

The Gemini API terms require users to be 18 or older, and forbid using the service in a site or app directed at, or likely to be used by, people under 18. In our reading, that rules out building a music generator for a children's app on this API. It is our interpretation, not legal advice.

On data, the terms separate unpaid from paid services. With unpaid services, such as AI Studio and free API quota, Google may use prompts and outputs to improve its products, and human reviewers may read them. With paid services, Google does not use prompts or responses to improve its products and keeps them only for a limited time for abuse monitoring. Since Lyria has no free tier, every API call is already paid. We still generate through the API with a billing-linked project rather than in AI Studio, so that commercial lyrics stay under the paid terms.

## Putting generated music on a video

We use Lyria in a small internal tool that scores short videos. The flow is generic: the tool samples a few frames from the video, a multimodal Gemini model proposes a genre, mood, tempo and instruments, the user adjusts them, Lyria generates the track, and ffmpeg replaces the original audio. Videos up to 30 seconds use the clip model; longer ones use the full-song model. If the music is shorter than the video it loops; if longer it is cut to the video's length with a fade at the end.

Two lessons came from it. Our users filled in the fields in Romanian, and since the current guide says Lyria writes lyrics in the language of the prompt, we translate the prompt to English before sending it. The first version of the translation step dropped the words *instrumental, no vocals*, which would have put singing over the video. We now add them back after translation if they are missing.

## Where we were wrong

**1. The list price as a cost estimate.** We first budgeted vocal songs at the list price of $0.08. Without iterative editing, and with silent lyric changes, the real cost per usable vocal track was about ten times that. The list price only holds for instrumentals that work on the first take.

**2. The dance name as the prompt.** We assumed that writing sârbă, horă or lăutăresc in the prompt was enough. It was not: sârbă came out as horă twice out of two. Musical description of tempo, accompaniment and structure is what the model can follow.

**3. Phonetic spelling as the fix.** Our working notes recommended respelling Romanian lyrics phonetically to fix diction. In the A/B, plain diacritics did just as well. Respelling is now a fix for specific words, not the default.

**4. Our own dance notes.** Our first wording for the sârbă specified a 2/4 meter, and our listener described the horă as three steps forward and three back. Published descriptions give other meters and step patterns. We removed both details from the prompts rather than guess.

**5. AI Studio as the data risk.** Our notes said to avoid AI Studio because human reviewers can read output. That is true for unpaid services in general, but the terms also say that AI Studio counts as a paid service when the account is linked to an active billing project, and that in the EEA, Switzerland and the UK the paid-service data terms apply to all services. For a Romanian company the gap is smaller than we wrote.

**6. Audio format.** Google's March 2026 changelog describes Lyria 3 output as 48 kHz stereo. Every file we received was a 44.1 kHz stereo MP3, which matches the current guide.

## Method

We generated tracks with `lyria-3-pro-preview` and `lyria-3-clip-preview` through the paid Gemini API between July and August 2026: functional instrumentals (lofi, sleep, ambient) and Romanian party genres (sârbă, horă, lăutăresc, manea, folk song with strigături, ballad), some with Romanian lyrics written by us. One listener familiar with Romanian folk and party music judged each track. Lyric fidelity was counted by comparing the sung words with the submitted text. Audio format and duration were read from the files with ffprobe.

We checked every platform fact against Google's own pages on 19 October 2026: the music generation guide, the Lyria prompt guide, the pricing page, the models page, the deprecations page, the changelog, the rate limit page, the Gemini API additional terms, and the Google Cloud prompting guide for Lyria 3 that lists the vocal languages.

We could not verify, and therefore dropped or flagged: a specific rate limit for Lyria, the mechanism behind silent lyric replacement, and claims in our notes about IP indemnification and ongoing litigation, which are outside this measurement.

## Limitations

The samples are small: eight functional tracks, two sârbă attempts and one or two tracks per other genre. The quality judgements come from one listener and are subjective. The 85% fidelity figure comes from our Romanian lyric tests, not from a large corpus. The diacritics versus phonetic A/B compared one track each, in two different genres, so it can only show that neither method failed. The prompt wordings for sârbă, horă and lăutăresc have not yet been tested in a second batch. All tests ran on Lyria 3, before Lyria 3.5 was released; the newer model may handle Romanian and its dances differently. Preview models can change without notice, and the cost multiplier depends on how strict the listener is.

## Sources

1. [Gemini API: Generate music with Lyria 3.5](https://ai.google.dev/gemini-api/docs/music-generation)
2. [Gemini API: Lyria prompt guide](https://ai.google.dev/gemini-api/docs/lyria-prompt-guide)
3. [Gemini API: Pricing](https://ai.google.dev/gemini-api/docs/pricing)
4. [Gemini API: Models](https://ai.google.dev/gemini-api/docs/models)
5. [Gemini API: Deprecations](https://ai.google.dev/gemini-api/docs/deprecations)
6. [Gemini API: Release notes (changelog)](https://ai.google.dev/gemini-api/docs/changelog)
7. [Gemini API: Rate limits](https://ai.google.dev/gemini-api/docs/rate-limits)
8. [Gemini API Additional Terms of Service](https://ai.google.dev/gemini-api/terms)
9. [Google Cloud Blog: Ultimate prompting guide for Lyria 3 Pro](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-lyria-3-pro)

## Questions and answers

### Does Google Lyria support Romanian lyrics?

Not officially. Google lists eight vocal languages for Lyria 3: English, German, Spanish, French, Hindi, Japanese, Korean and Portuguese. In our tests Romanian worked anyway: about 85% of the words were sung as written, diacritics were pronounced correctly, and the intonation was odd in places. It is not guaranteed and may change.

### Should I write Romanian lyrics with diacritics or phonetically?

In our A/B test neither won. Plain Romanian with diacritics worked as well as an English-style phonetic respelling. Start with diacritics, put the lyrics after a Lyrics: header, keep blocks to four lines, and respell only the words that come out wrong.

### How do I get a real sârbă or horă from Lyria?

Describe the music, not the dance name. For a sârbă, ask for a very fast driving dance around 150 BPM with relentless off-beat oom-pah accompaniment and a continuous rapid melody. For a horă, ask for a moderate circle dance around 100 BPM with a cyclical, swaying phrase that rocks forward and back. Named alone, sârbă came out as horă in both our attempts.

### How much does Lyria cost per song?

The list price is $0.04 per 30-second clip and $0.08 per full song for Lyria 3, and $0.08 per song for Lyria 3.5, with no free tier. For vocal songs with fixed lyrics, expect the real cost of one usable track to be about ten times that, because each retry is a full new generation.

### Is Lyria music watermarked?

Yes. Google states that all generated audio includes a SynthID audio watermark, and its Cloud blog says outputs also support C2PA content credentials.

### Can I use the Gemini API to make music in an app for teenagers?

The Gemini API terms say users must be 18 or older and that the service must not be used in a site or app directed at, or likely to be used by, people under 18. In our reading that excludes apps aimed at teenagers, but this is an interpretation, not legal advice.

## How to cite

**Cite this as:** Going Stellar (2026). *Can Google Lyria 3 sing in Romanian and play Romanian genres? What we measured through the Gemini API*. GOING STELLAR S.R.L.. https://goingstellar.ro/research/google-lyria-3-romanian-music-test/

```bibtex
@misc{goingstellar2026google,
  title  = {Can Google Lyria 3 sing in Romanian and play Romanian genres? What we measured through the Gemini API},
  author = {{Going Stellar}},
  year   = {2026},
  url    = {https://goingstellar.ro/research/google-lyria-3-romanian-music-test/},
  note   = {2026-10-19}
}
```

Licence. Free to quote with attribution.

## Also available

- This page as HTML: https://goingstellar.ro/research/google-lyria-3-romanian-music-test/
- Same article in Romanian: https://goingstellar.ro/ro/cercetare/google-lyria-3-muzica-romaneasca-test/