What merging RVC models actually does
An RVC voice model is a file of learned weights, usually a .pth, that describes how one voice sounds. Merging two models means mixing those weights at a ratio you choose, then saving the result as a new model. Run a vocal through it and you hear a voice that sits between the two parents.
At 50/50 you get a true middle. At 70/30 the result leans toward the first voice but picks up the color of the second. Small shifts change a lot, so it is worth trying a few ratios and listening to each.
The new voice does not sound like either parent, which is the point. You end up with an original voice that belongs to nobody else, useful for an artist project, a character, or a voice you can release without imitating a real singer.
Which models can be merged
Only models with the same shape can be mixed. In practice that means the same RVC generation, v1 with v1 or v2 with v2, and the same sample rate. Try to merge a v1 and a v2 model and the weights simply do not line up.
Models you train yourself in the same app with the same settings always match. Models downloaded from different places vary, so check what each one is before you try. A quick way to tell is the file's details in whatever tool you use; good tools show the generation for you.
The do-it-yourself way
The original RVC WebUI and several community forks include a model merge tab. You point it at two .pth files, set the weight, choose a few options and export. It works, but you need the WebUI running, which means Python, dependencies and a fair amount of setup, and it will not always explain why a pair failed.
If you go this route, keep backups of both parent models and write down the ratio you used for each merge, because the file name is all you have later.
Merging vs training a new voice
Training builds a voice from recordings: you collect minutes of clean audio from one person, train for an hour or so, and get a model that sounds like them. Merging builds a voice from voices: no recordings, no training time, and the result sounds like no one in particular.
That makes them good at different jobs. Train when you want a specific voice, such as your own, so you can type lines or fix a take without re-recording. Merge when you want something new: an original artist voice, a character, or a variation on a voice you already like.
The two work well together. Train a clone of your own voice with voice cloning, then blend it with a starter voice at 70/30. You keep your phrasing and character while the tone shifts toward the other voice, which is a quick way to try a different sound without losing what makes it yours.
The one-click way in Xport Studio
The Voice Blender in Xport Studio does the same merge from the Library. Pick Voice A and Voice B, drag the slider from 5 to 95 percent, and press Blend. The new voice lands in your Library with a name like Nova x Iris 50/50, ready for the vocal converter or text to speech.
Each voice shows an RVC v1 or v2 badge, and when a pair cannot mix the app tells you exactly why instead of failing with a generic error. Every voice you train inside Xport matches every other trained one, so blending your own voices always works. It runs on your Mac with nothing uploaded.
Tips for better blends
Start from two clean models. A merge cannot fix a muddy or noisy parent, it passes the problem on. Voices in a similar range blend more naturally than a very low voice with a very high one, though that pairing can make interesting characters.
Listen to each blend on the same test line, ideally the kind of material you will actually use it for. If you plan to sing with it, test with a sung take; for voiceover, use speech. And only merge voices you have the right to use: your own, starter voices, or models licensed to you.