Your reusable assets are the kindling. The scene is how you lay the fire. Rendering is the burn.
Lay kindling: upload your audio
Pick a recording from your files. It is analyzed in your browser, nothing leaves your device.
Your kindling
Saved voice characters, instruments, and presets. Stored on this device.
Calibrate words for
Some words, the reading rules are likely to get wrong for this voice. Give one take of each word you care about: sing or say the word once, on its own, close to the mic in a quiet, dry room (no reverb, no effects). Singing or speaking both work, because the fit measures the shape of the vowel (its formants), and formants do not depend on pitch. One good take per word is the design point.
Each accepted take is measured and stored as numbers inside this voice profile: a verified reading for that word, small corrections that also help other words sharing the same sounds, and the take itself, kept with its date and source under the same consent record as the voice. This is a calibration layer stored as data in the profile. It is not a trained model. A take that is clipped, too short, or too noisy is refused with the reason shown, and nothing is fitted from it.
Burn: compose and render
0:00 / 0:00
VoiceInstrDrums
Voice
Instr
Drums
Master
Session setup
What changed
Every update, newest first. Nothing is ever removed, each version stays in the record.