After training

How to use your RVC model after training

Training gives you files, not a finished sound. This guide covers where the .pth and .index belong, how to load them in the interfaces people actually use, and what to check when a model does not appear in the list.

What you downloaded

A completed run produces a model-weight file and a retrieval index. The .pth holds the trained weights and is the file an RVC interface loads as the voice. The .index is built from your prepared audio and is read during conversion. The ZIP simply keeps the matching pair together under your model name. See .pth vs .index for the full comparison.

Where each file goes

Folder names differ between RVC interfaces, but the pattern is consistent: model weights go in a weights or models directory, and the index goes in the index or logs directory. After copying files in, use the interface's refresh control rather than restarting — most builds only rescan those folders on demand.

Do not use the checkpoint files

Local RVC training also writes files beginning with D_ and G_. Those are training checkpoints, not loadable voice models, and selecting one is a common reason conversion fails. Your NiceVois download is the final model, so this only matters if you are mixing in files from a local run. The optional early checkpoint offered during training is a validated, usable model and is named accordingly.

Running your first conversion

Select your model, load the audio you want converted, and convert. Two settings account for most poor first results:

• Pitch. Shift by -12 or +12 semitones when the source and target voices sit an octave apart, typically when converting between a lower and a higher voice. Leave it at 0 when they are in a similar range.
• Index ratio. Higher values lean harder on the retrieval index and sound closer to the training data; lower values track the input more literally. Mid-range is a reasonable starting point.

Making a song cover

Conversion works on a vocal, not a finished mix. Separate the vocal from the instrumental first, convert the isolated vocal with your model, then place the converted vocal back over the original instrumental. Feeding a full mix straight in is the single most common cause of a muddy result.

Keep the pair together

An index belongs to the run that produced it. If you rename, move, or archive the model later, keep the .pth and its matching .index together and preserve the names, or the index will no longer be found.