Free browser tool
Check and merge RVC voice models in your browser.
Inspect the technical values of one RVC checkpoint, or choose two models to see exactly where they match. Compatible pairs can be blended into a new .pth model without leaving your device.
How it works
Create a blended RVC checkpoint in four steps
The merger blends the trained tensor weights inside two compatible RVC checkpoints. It does not generate audio, retrain either model, or send the models to a server.
1. Choose two .pth files
Download the two RVC checkpoints you want to combine, then select both files above.
2. Check compatibility
The browser reads their architecture and enables merging only when the important checkpoint details match.
3. Set the blend
Choose the contribution from Model A and Model B, and select which checkpoint supplies the saved metadata.
4. Download and test
Create the merged checkpoint locally, save the optional recipe, and test the new voice in a compatible RVC app.
Which RVC models can be merged?
Two files both being called “RVC” is not enough. They need matching checkpoint structures: RVC version, sample rate, pitch mode, vocoder, tensor types, tensor names, and tensor shapes. The compatibility check runs before the merge and fails closed when those details do not line up.
Compatibility only confirms that the files can be blended safely. It cannot predict whether the two voices will make an appealing combination.
What stays on your device?
Everything involved in the merge: both source checkpoints, checkpoint inspection, tensor blending, the finished model, and the optional recipe. The tool makes no model-upload request to AIVoices, Supabase, or another processing service.
Large checkpoints still need substantial browser memory. The tool estimates the requirement and blocks a pair when the device does not report enough capacity.
Find models to blend
Start with two RVC checkpoints
Browse the RVC directory to find voices, then download the exact .pth checkpoints from their original sources. Catalogue-assisted compatibility matching and category-based pair discovery are planned for a later release; the merger itself works now with local files.
RVC model merger FAQ
Questions about merging RVC models
How can I check the values of an RVC model?
Select one .pth checkpoint in either model slot. The browser will show its RVC version, sample rate, pitch mode, vocoder, checkpoint layout, tensor format, tensor count, merge status, and technical fingerprint.
Can I merge any two RVC models?
No. The checkpoints need matching RVC architecture details, including version, sample rate, pitch mode, vocoder, tensor data types, keys, and shapes. The tool checks these details before enabling the merge.
Are my RVC models uploaded to AIVoices?
No. File inspection, compatibility checks, weight blending, and output creation run in your browser. Your checkpoints are not sent to AIVoices.
What does the blend percentage change?
It controls how much each compatible checkpoint contributes to the merged tensor weights. A 50/50 blend uses equal weight from both models; moving the slider favors one model.
What happens to the .index files?
The merger only creates a new .pth checkpoint. It does not merge or include index files. Test the merged checkpoint without an index first, because an index trained for either original voice may not suit the blend.
Does a compatible merge guarantee a good voice?
No. Compatibility means the checkpoint structures can be blended safely. The result may sound useful, subtle, unusual, or poor depending on the two voices and their training.
Where can I use the merged model?
Use it in an RVC application that supports the checkpoint version and architecture shown by the compatibility check. Keep both original models and test the blend before using it in a project.
Before you use a merged voice
Keep the original checkpoints, listen to the result, and follow the terms attached to both source models. AIVoices provides the local merging tool but does not grant rights to a model, dataset, character, performer, or generated output.