RVC model training learns the characteristics of a voice from a dataset and produces portable model files. NiceVois handles the cloud environment, preprocessing, training, validation, and packaging so you do not need a local NVIDIA GPU or a notebook session.
1. Start with audio you have permission to use
Use your own voice or audio from a speaker who has knowingly given you permission to train a reusable voice model. Permission to possess or publish a recording does not automatically include permission to create a model from it.
2. Prepare a clean recording
Choose a supported WAV, MP3, FLAC, M4A, or OGG file. NiceVois shows an exact workload quote before training, and your first training is free with a free account. Clear, dry, single-speaker audio is more valuable than a longer recording with music, reverb, clipping, or competing voices.
Practical rule: recordings shorter than one minute may not provide enough voice variety for a reliable model. Use as much clean material as the displayed limit permits.
3. Name the finished model
The model name becomes the filename for the downloadable .pth, .index, and complete ZIP package. Choose a short descriptive name without file-path symbols.
4. Start with the recommended epoch count
An epoch is one pass through the prepared training dataset. The interface reads the audio duration and selects a practical starting recommendation. More epochs take longer and are not automatically better; excessive training can overfit the source material.
5. Follow the real training stages
After the upload completes, the server owns the job. The page reports audio receipt, dataset preparation, GPU assignment, live completed epochs, model validation, and final packaging. You can close the browser or shut down your computer once training has started.
6. Download the output you need
- .pth: the trained RVC model weights.
- .index: the retrieval index built from the extracted voice features.
- ZIP package: the complete portable bundle containing both.
When training has produced a validated early checkpoint, the workspace can also expose that .pth while the final run continues.
After training: hear the model work
Training is only half the job - a model you never run is a file you cannot judge. Convert a vocal or a whole song with the finished voice in the RVC voice converter, which returns the normal RVC result alongside a Clean Consonants version. To put your voice on a finished song, the Suno voice changer separates the track, converts the lead, and mixes it back.
Open the converter, upload a supported recording, and follow the duration-based recommendation.
Train an RVC model from audio