Lyria prompt section markers: [0:00-0:30] Intro on Sume Music

Sume Music takes its structure from the prompt. There is no duration or duration_seconds field; the docs say the API rejects both. To get a structured track, write the length in words ("a 2-minute track") and add section markers in the form [0:00-0:30] Intro: .... Music 1.0 runs on Google Lyria 3.5 and makes full-length structured songs of up to a few minutes, so the markers are how you divide that time.

A marked-up brief

The marker syntax comes straight from the Sume docs. The musical axes below are the docs' seven-axis brief: emotion, genre, tempo as a number, key and mode, two to four textured instruments, an arc with one named moment, and era or production. They are creative directions, not guaranteed output values, so listen to the result.

A 1-minute track. 96 BPM, A minor, warm analog synthwave.
[0:00-0:15] Intro: soft pad and a muted bass pulse, sparse.
[0:15-0:40] Verse: add gated drums and a plucked lead, restless.
[0:40-1:00] Outro: strip back to pad, slow fade.
Dry and close, 1986 production. Instrumental, no vocals.

What the API does with it

Nothing special. The prompt is sent as text (1 to 5000 characters), and the accepted request is charged a fixed price. The markers never change the price, which stays the same for a 30-second brief and a 3-minute one.

Why lyrics output matters

When the provider returns them, result.lyrics carries model-reported lyrics or a section map. The docs call this metadata, not an audio measurement. Do not use it to prove the track matches your markers; compare by ear, or measure the audio artifact yourself.

Retry rules

If a policy rejection comes back, change the flagged content but keep the musical brief, and retry only within the budget you were given. Do not shrink the request to a generic bed, because the brief is the part you pay for. Use an Idempotency-Key so a network retry does not create a second $0.125 job.

A checklist before you submit

Check three things. First, the marker times should add up to the length you wrote in the sentence at the top of the prompt, so a "1-minute track" ends its last marker at 1:00. Second, keep each section to one clear change, since the model reads the line as direction, not as a score. Third, end with the instrumental clause so vocals do not appear in a bed.

• Wrap the request in an Idempotency-Key.
• Poll GET /v1/jobs/:id/status, then read result.artifacts[] for the audio entry.
• Listen to the result and compare the section changes by ear.