Model training here means fine-tuning an existing open model on your own data rather than training one from nothing. NUZM has done this most often with XTTS for voice cloning, and the same method applies to any task where a general model is close but not specific enough to your material.
What this includes
- An honest assessment of whether your data is enough, before any GPU time is billed.
- Dataset preparation and cleaning, which is most of the work and most of the result.
- A held-out set the model never sees during training, so the evaluation means something.
- The trained weights handed to you, so you are not renting access to your own model.
- A written note of what it does badly, which is as useful as knowing what it does well.
Is this the right choice for you?
Right when a general model almost works: it handles your domain roughly, gets your terminology wrong, or does not sound like the voice you need.
Wrong when prompting would have done. Fine-tuning costs money, takes time and produces something you then have to host; a well-written prompt against a hosted model costs neither and is often indistinguishable in the result.
Worth knowing before you commit
Voice cloning has a consent problem before it has a technical one. Cloning a voice that is not yours, or that its owner has not agreed to, is a legal exposure that no amount of model quality fixes. We ask who the voice belongs to before we ask how many hours of it there are.