Artist voices (LoRA)
Train an ACE-Step voice LoRA for any act in the WeavyAudio catalogue and use it for new songs, covers and edits.
A voice is an ACE-Step LoRA trained on every published song of one act, plus one of that act's records as the reference voice. In the studio the voice list sits under the engine picker when ACE-Step is selected.
Train a voice
Open Voices, search the act and press Train voice. The studio shows an estimate before it starts – roughly 40 seconds of A100 time per song (40 epochs, rank 32) plus about 10 minutes of setup. An act needs at least 5 published songs. One voice trains at a time; the list updates live and the voice becomes selectable when it is ready.
How it is used
Each ACE-Step job loads the voice's LoRA on the ACE server, renders, and puts the house sound back – the adapter is global on that server, so ACE-Step jobs run one after another. The reference record is sent with every job unless you choose Reference voice with your own source.