Same Lyria 3.5 prompt, different song: why, and no seed on Sume
Google says Lyria results vary between calls. Sume's music API has no seed. How to keep the take you like and generate a few to choose from.
The same Lyria 3.5 prompt can give a different song every time, and Sume's music docs list no seed to fix it. Google's music-generation guide, read 2026-10-01, says results vary between calls and that generation is single-turn, with no multi-turn editing.
What should I do instead?
Treat each call as a take. Generate two or three with different Idempotency-Keys, listen, and keep the artifact you like. Each costs a fixed $0.125, and the saved file on media.sume.com is your master.
| Item | Value |
|---|---|
| Seed | None listed |
| Cost per take | $0.125 fixed |
| Edit call | None; change the prompt and generate again |
| Retry with the same Idempotency-Key | Returns the same job |
| New Idempotency-Key | A new take |
Can I tweak a track?
Not in place. There is no edit call; change the prompt and generate again. To trim or join, use Timeline audio.
What about retries?
An Idempotency-Key makes a retry return the same job rather than a new song. A new key is a new take.
Sources
Related posts
More in Developers
- MAI-Image-2.6-Flash API: not on Sume, models to use instead
Microsoft's MAI-Image-2.6 and 2.6-Flash launched 2026-09-04 in Foundry preview. Sume's image catalog does not list them; check GET /v1/images/models.
- MAI-Transcribe-2 word timestamps vs Sume STT words[] shape
MAI-Transcribe-2 gives every word a start and end time. Sume STT always returns words[] as { word, start, end } in seconds, with no flag to turn it on.
- MAI-Voice-2.1: 23 languages and voice matching; Sume takes a voice id
Microsoft says MAI-Voice-2.1 covers 23 languages and matches a voice from a short clip. Sume TTS takes a ready voice id or avatar, not reference audio.
- MAI-Voice-2.1-Flash lists 45 ms; Sume TTS is an async job
Microsoft lists MAI-Voice-2.1-Flash at about 45 ms and MAI-Voice-2.1 at about 550 ms. Sume TTS 1.0 is non-streaming: a job with a poll URL or webhook.
Written by Sume