Skip to main content
You can upload a vocal recording or save vocals from an owned song. Saving from a song first extracts its vocals. The saved profile keeps the sample and preparation status.

​
Reference conditioning

Reference mode conditions generation on the recording. A saved sample_url makes that recording available before training finishes. You can also submit a direct voiceReferenceUrl instead of a saved voice ID. The resemblance control maps to voiceCoverStrength. It ranges from 0 to 1 in the API. Voice integration covers the request fields.

​
Train for reuse

The saved-voice flow prepares LoRA adapters for the configured Turbo and SFT checkpoints. LoRA trains a small adaptation for the selected model while keeping the underlying checkpoint reusable. Preparation uses isolated mono or stereo vocals up to four minutes. Each model has its own availability status, so the same profile can be ready for one mode while another is still processing. Training uses its own worker pool. Inference applies the selected adapter under an exclusive GPU gate, then disables or restores it after the job. The hot cache reuses a loaded adapter when the same voice returns.

​
Use the result

For trained mode, check voice.availability[modelVariant].status before generation. Send voiceId and voiceMode: "trained". For a recording, use voiceMode: "reference" with a saved ID or direct URL. The voice walkthrough shows selection and resemblance. Voice events report preparation without repeated submissions.