Suno v5.5 (March 2026) did not rewrite Style syntax—it added three personalization layers. Mixing them blindly is why “the same prompt sounds wrong today.”

What each layer controls

Layer	Controls	Who	Leave off when
Voices	Whose vocal timbre sings	Pro/Premier (verified capture/upload)	You need a generic demo singer
Custom Models	Your catalog’s arrangement DNA	Pro/Premier (train on owned tracks)	Training set is stylistically chaotic
My Taste	Defaults / Magic Wand lean	All users	You are A/B testing Style strings

Style field still owns ~first-listen genre. Personalization biases; it does not replace tags.

Voices — identity, not karaoke magic

Capture or upload your singing (clean acapella beats mixed beds). Use Voices when the release must sound like you. Do not expect it to fix weak lyrics structure—keep [Verse] / [Chorus] tags honest.

Open Suno in MidassAI

Custom Models — style from your catalog

Upload a tight set of tracks you own (official guidance centers on a small minimum catalog—think focused, not “everything you ever made”). Up to a few models per account is typical; train one lane per sound (lo-fi bedroom vs live band).

If the training mix spans five genres, the custom model will average into mush—same failure mode as overstuffed Style tags.

My Taste — quiet and always-on-ish

My Taste learns from what you play and make, then nudges suggestions (Magic Wand / style augmentation). When debugging prompts, disable taste augmentation and compare—otherwise you will “fix” Style strings that were never the bug.

Recommended onboarding order

Lock a clean Style stack on stock v5.5 (see our style prompt guide).
Enable My Taste only after you like the baseline.
Add Voices for identity releases.
Train Custom Models last, once you have a coherent catalog slice.

Prerequisites and setup

Select model v5.5 explicitly. Custom Mode required for serious Style/Lyrics control. Confirm commercial-rights tier before client delivery—plan names and prices change; check Suno’s current terms.

Extended workflow

Generate two takes with personalization off.
Re-run identical Style + Lyrics with My Taste on; note only default drift.
Add Voices on the keeper if identity matters.
Only then swap in a Custom Model and regenerate once.
Save the winning combo as a MidassAI preset (model + voice + custom + style string).

Common mistakes

Training Custom Models on covers you do not own. Fix: rights-clear originals only.
Using Voices to hide bad section tags. Fix: repair structure first.
Chasing prompt bugs with Taste on. Fix: baseline off, then re-enable.

Try this in MidassAI

Keep personalization toggles beside the Style preset so collaborators see the full recipe. Open https://www.midassai.com/studio/suno/ and A/B one chorus with Taste off vs on before you touch Custom Models.

Suno v5.5 Voices, Custom Models & My Taste Explained