Harmonic tendencies
Which chord movements and tonal centres recur in your work.
Transparency
How training on your own music works
No mystique and no overclaiming: here is what happens when you train a Sound, what the model picks up from your material, and where its understanding ends.
Process
From files to a usable model
Selected filesPrepared datasetTraining runSound version
Preparation matters as much as training. Material recorded across years at different levels and lengths is normalised and trimmed, and converted into the representations the training run consumes, so the model learns from your music rather than from inconsistent file handling.
The run itself takes your configuration — what it should focus on, the base setup and basic hyperparameters such as how many passes to make over the data and how aggressively to learn — and produces a versioned Sound recorded with the material used, time taken and resulting size.
Signal
What the model learns
Melodic contour
Interval habits and phrase shapes typical of your writing.
Rhythmic feel
Groove, subdivision and the tempo territory you actually work in.
Timbre
Instrumentation and sonic texture present in the recordings you provided.
Production character
How your material is typically mixed and layered.
Structure
How your songs tend to move between sections and manage energy.
Boundaries
What it does not learn
A Sound has no model of your intent, your lyrics' meaning, your audience or your career. It does not memorise your catalog for playback, it will not reliably reproduce a specific performance, and it is not a voice clone — vocal character may influence output where you trained on vocal material, but treating it as a controllable voice replica is misunderstanding the tool.
It also does not have taste. Generation produces candidates; whether any of them is worth pursuing is a judgement only you can make. That is why the workflow ends in Directions rather than in a finished master.
Prerequisites
Data and rights requirements
Data
- Audio you own or control — mixes, demos, stems, vocals, instrumentals
- Higher-quality sources give preparation more to work with
- Consistent material produces a more coherent Sound
- Optional MIDI and lyrics for related workflows
Rights
- Composition, recording and performance rights for the material
- Clearance for any third-party samples it contains
- Agreement from co-writers, featured artists or rights holders
- A separate confirmation for voice recordings before training
Handling
Privacy during training
Training reads only your dataset and writes a Sound owned by your account. The audio required for the run is processed by connected specialist compute configured for your studio; nothing is added to a shared model or exposed to other accounts. The privacy for artists page covers processors, retention and deletion specifically.
FAQ
Frequently asked questions
Related
Keep reading
Train a Sound you understand.
Clear inputs, honest limits, and a model that belongs to your account.