Spanish song lyrics from an AI music prompt: write it in Spanish

Write the whole prompt in Spanish, including any lyrics you supply, because Google's Lyria documentation says the lyrics follow the language of the prompt. On Sume, send that prompt to the Music Router with lyria-3.5 or sume/music-auto, then check the model-reported section map in the result.

There is no language switch to flip. The language is a property of the text you send, so a Spanish brief with English genre words can drift, and a fully Spanish brief is the safest way to ask for Spanish vocals.

What Google documents

The Gemini API music generation page says Lyria 3.5 supports custom lyrics with section tags such as [Verse] and [Chorus], and that generated lyrics follow the prompt language. It also states that requests for an artist's voice and for copyrighted lyrics are blocked, and that the interface is single-turn only. Output is MP3 or WAV at 44.1 kHz stereo and carries a SynthID watermark.

Single-turn means you cannot reply with a correction to the same session. To change a line, edit the prompt and generate again.

A prompt shape that works

Keep section tags in the same bracketed form Google documents, and keep the instructions in Spanish too, so the model does not see mixed signals about language. A skeleton looks like this in plain text: a short description of style and mood, followed by [Verse] with two or four lines, [Chorus] with the hook, and a closing [Verse] or [Outro].

• Describe tempo feel, instruments and mood in Spanish.
• Write original lyrics rather than quoting a song; copyrighted lyrics are blocked.
• Keep each section short, since the Sume router rejects duration fields and the length is the model's choice.
• Use the optional image_url only when a still really sets the scene.

Checking the result on Sume

The Music Router docs say a music result can carry result.lyrics, a model-reported section map. Compare it with the lyrics you sent: if a section is missing or reordered, the track did not follow your structure and a re-run with a tighter prompt is the fix. Treat the map as the model's own report, not as a forced alignment. The lyrics section map post explains the difference.

Listen for pronunciation on names and borrowed words. Because there is no per-word control, the usual correction is to respell the word phonetically in the prompt.

Rights and labels

Lyria output carries a SynthID watermark, and Google blocks artist-voice requests. If you publish on YouTube, check whether the music is the main focus of the video, since the rules differ by use; see our YouTube label post. The same prompt discipline applies to any language: say what you want in the language you want it sung.