Complete workflow

How to train an RVC voice model online

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. The current beta accepts up to five minutes. Clear, dry, single-speaker audio is more valuable than a longer recording with music, reverb, clipping, or competing voices.

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.