What a trained RVC model is
RVC stands for Retrieval-based Voice Conversion. Training it means teaching a model how one specific voice sounds, so it can later take any vocal and re-sing or re-speak it in that voice. The result is two files: a small .pth file of learned weights, usually around 60 MB, and an optional .index file that helps the model stay close to the voice it learned.
The model does not learn words or melodies. It learns tone, texture, and character. That is why the quality of what you feed it matters far more than any setting you touch later.
How many minutes of vocals you need
The RVC project's own FAQ recommends around 10 to 50 minutes of audio. It also says 5 to 10 minutes can be enough when the voice is very distinctive and the recording is clean, and that results from under a minute have not been reproducible. More audio only helps if it stays clean and consistent: one singer, one tone, low background noise.
Quality beats length every time. Ten minutes of dry, well-recorded vocals will train a better model than an hour of live recordings with crowd noise and reverb. If your best takes are buried in a mix, pull them out first with the stem splitter, then clean up hiss or hum with the noise reducer.
Preparing the dataset
Use WAV files at 44.1 or 48 kHz. Cut out long silences, coughs, other people talking, and backing vocals that are not the target voice. Harmonies and doubles sung by the same person are fine, but a second singer will blur the model. The audio cutter handles quick trims in the browser.
Keep the recording chain consistent if you can: same microphone, same room, similar distance. A model trained on three very different mic setups learns an average of all three. And avoid spaces, brackets, and non-English characters in file paths; the RVC FAQ lists path characters as a common cause of ffmpeg and encoding errors during training.
How many epochs to train
An epoch is one full pass over your dataset. The RVC FAQ is direct about this: if your audio is noisy, 20 to 30 epochs is enough, and more will not improve it. If your audio is clean, low-noise, and long enough, you can go higher, and around 200 is described as acceptable.
Training longer is not automatically better. Past a point the model starts copying flaws in the dataset, which shows up as harsh, metallic, or over-sharp results. Save checkpoints along the way and compare them on the same test line. Often an earlier checkpoint sounds more natural than the last one.
What the index file does
RVC starts from a pretrained base model, so a little of that base can leak into your results. The index file is a lookup of features from your training audio, and the index rate setting controls how strongly the model pulls toward it. Higher values keep the tone closer to your training voice; lower values can sound cleaner when the training audio itself was rough.
Always keep each .pth with the .index file made from the same training run. Mixing an index from one run with a model from another is a common cause of odd results. Do not change the sample rate halfway through a training run either; the FAQ notes that it breaks training and you have to start a new experiment.
Test before you trust it
Run the same short test line through every checkpoint you are considering: one sung phrase and one spoken sentence. Listen for breaths, s and t sounds, and the top of the range, which is where weak models fall apart first. Our guide to the best RVC settings covers the knobs to try before you decide a model is bad.
If the model only sounds right at one pitch, it probably needs more varied training audio rather than more epochs.
Training without Python
The RVC WebUI and its forks train well, but they expect Python, dependencies, a compatible GPU setup, and some patience with error messages. Train Clone in Xport Studio runs the same kind of RVC training as a desktop app. Add 5 to 30 minutes of clean vocals, press Start, and a typical run finishes in 30 to 90 minutes on your own computer, with nothing uploaded.
The finished voice lands in your Library, ready for the vocal converter or text to speech. Train Clone is part of Xport Pro, $79 one time, and runs on macOS 15 or later on Apple Silicon, or Windows 10 or later. More on how it works on the voice cloning page.