How Suno clips relate to each other

Every clip Suno generates carries metadata about where it came from. Building suno-web meant working out that structure from an archive of about 19,000 clips.

A clip can point at other clips in nine ways:

• cover_clip_id: it is a cover of that clip
• edited_clip_id: it is an edit
• artist_clip_id: it borrows that clip’s voice
• stem_from_id: it is a stem separated from it
• styles_lyrics_clip_id: it reuses the style and lyrics
• mashup_clip_ids: a list of clips it combines
• history: the chain of extensions, each with the time it continues from
• concat_history: the parts joined to make a full song
• persona.root_clip_id: the clip a persona was made from

The kind of generation is in metadata.task, which takes about 25 values.

For a single “parent” I take the first of cover, edit, artist, stem and styles that is set. Walking up is guarded against cycles. Walking down needs a budget: one family in my archive has 426 children.

Other things the data taught me:

• A persona’s name is inside the persona object. The top-level persona_name field is not where it lives.
• Suno returns two clips per generation with the same title and the same created_at to the millisecond, so a list has to group them or everything appears twice.
• There are no word timings. Lyrics can only scroll in proportion to playback position; real synchronisation would need forced alignment against the audio.