Keep AI Studio for custom renders, then move to the Store for catalog tracks, SFX, plugins, and licensing coverage on client work.
Grab a free starter kit — 50 sounds, no card.
Drum hits, one-shots, a few loops. Open in any DAW.A free-text prompt is good at genre and bad at timbre. A persona is a stable acoustic anchor — built from reference clips, an embedding, and a prompt prefix. Here is how the library is built and how to use it.
May 9, 2026 · 4 min read · SONICHAOS editorialA prompt tells the model what room to walk into. A persona tells it who is already in the room.
The first version of the SONICHAOS AI Studio composer shipped without a persona library. You typed a prompt, you got a song. The renders were fine on genre and tempo. They were inconsistent on timbre, vocal range, and accent — the qualities that make a voice recognisable. A second render of the same prompt could shift the singer from a tenor to an alto without warning. The fix was not a longer prompt. The fix was a named, stable voice on top of the prompt. That is the persona library.
A persona on SONICHAOS AI Studio is three artefacts, glued by a stable id:
WavLM
features and pooled over time. The pooled vector is the voice
fingerprint. Cosine distance between two pooled vectors lines up
well with how a listener would describe the difference.A persona is not a checkpoint, not a fine-tune, not a LoRA. It is data. Adding a persona does not retrain anything — it adds a row to the persona index and a folder of clips to object storage. New personas ship in a day, not a sprint.
The composer joins persona and prompt at the embedding stage, not at the text stage. That detail is load-bearing. If you put the persona description in front of the user's text, the model treats it like a request and tries to satisfy both. Two halves of the prompt fight each other and the result drifts.
Holding the persona prefix at the embedding layer fixes that. The voice is already chosen by the time the prompt is parsed. The free-text prompt then steers the song around that voice — genre, tempo, mood, lyric.
A few rules that fall out of this:
Glasspine persona selected pulls the model in two
directions. The persona wins, but the render takes a noticeable
quality hit.Three of the demo personas are worth picking up if you are sketching:
The other three personas — Arc Koi, Glasspine, Rico Lima — cover hooky tenor, choir-stack soprano, and Spanish-language flow at 86 to 96 BPM. The full list is on the personas page.
A community-uploaded persona system is on the roadmap. We built the fixed library first because the embedding pipeline, the prompt prefix contract, and the audit-log shape are easier to test against six known voices than against an open set. Every persona in the v1 library has been re-rendered against ten reference prompts, twice, and the cosine distance between the two renders is below our drift threshold. That sets the bar that user-uploaded personas will need to clear.
A persona is data. The library is small on purpose. It is a place to start.