Change one thing. Keep the song.
Sunofriend is planning a second in-development feature: identity-preserving remixing. A bounded remix makes one intentional change—one instrument, role or region—while keeping the motifs, harmony, groove and structure the owner recognises from the source. No remix model is installed, trained or authorised yet.
A canonical plan exists. The feature does not.
The public programme plan defines how remixing and audio-native vocal comping will be built together around a shared, time-aligned Musical State. That plan authorises no model installation, training, cloud upload or production feature by itself.
There is no remix command, no trained model and no generation yet. The first remix controls are planned as deterministic assembly or retrieval with frozen pretrained models (D+I+H)—no training required. Learned conditioning (T+I+H) would begin only after repeated bounded remix evidence beats those controls on real songs.
What counts as a remix here.
Not unconstrained full-song regeneration. The owner names what must stay; the operation changes only what was permitted.
Motifs
Recognisable accompaniment and melody fragments the owner can name.
Bass and harmony motion
The harmonic movement and low-end direction that carry the song.
Groove and section energy
How drums, feel and arrangement energy change between sections.
Lyrics, phrase and structure
Canonical words, phrase boundaries and the song's shape.
- One permitted change. A remix benchmark names the identity anchor and the single thing allowed to change—an instrument, a role, a region—before anything is rendered.
- Bounded first. The first planned operations cover 8–16 bars, not a whole song.
- Deterministic control first. Every learned or generated attempt is compared with a deterministic control, and fails if every owner-recognised anchor is lost.
- Editable handoff. Reviewed results stay editable—MIDI/Clip assembly or region-level edits—and keep an exact source map.
- Fixture-specific identity. Which elements carry identity is decided per track by listening; it must not become a universal rule.
How the work is labelled.
Every task and experiment declares what kind of work it is, so a trained model can never quietly stand in for human judgement.
Deterministic
Fixed code and rules edit audio, timing, manifests or edit maps. No learned weights are consulted or changed.
Frozen-model inference
Existing pretrained models are used for analysis or generation. The model is used, not trained.
Model training
An optimisation job changes learned weights. Every trained output stays a research challenger until real-song evidence earns promotion.
Human musical review
The musician listens, chooses, rejects and names what must stay. This is the musical authority.
What the design must protect.
- Your audio stays local. The public site has no upload or hosted remix endpoint.
- Listening outranks similarity scores. Embedding similarity and MIDI/F0 self-agreement do not establish the feature.
- Training is gated. Local experiments may begin early, but every checkpoint is a challenger; larger or paid cloud training needs separate authorisation and never uploads private audio.
- Generated audio is a labelled source class. It carries consent, training provenance and visible edit-map labelling; a human/AI duet is valid only when chosen by the user and is never described as fully human.
- No acceptable result is a valid result. If neither the deterministic control nor any challenger keeps the named anchors, the failure is retained and the next cycle is chosen.
One bounded remix on one authorised track.
The planned first cycle names a musical identity anchor and the one permitted change, renders a deterministic control, adds a frozen-model challenger only where it helps, and plays both against the track in a real music session. The accepted result—or the exact reason neither is useful—becomes the evidence for the next cycle.
Development alternates between vocal comping and remixing on real songs, so both features are measured by playable artifacts, not metrics alone.