Dataset
Why denoising your training data makes a better voice model
A voice model learns the statistics of the file, not just the voice in it. A constant hiss, a hum, or the sound of a room is present in every single example, so the model treats it as part of the singer and reproduces it in every conversion afterwards. Cleaning the dataset removes it once. Cleaning the output means fighting it forever.
This is the highest-leverage cleanup you can do, and it is the one people skip — because the noise is quiet, and because it does not sound like a problem until it is baked into a model.
Why the model cannot tell the difference
Training does not separate “the voice” from “the recording of the voice”. It looks at your clips and learns what they have in common. If every clip has the same air-conditioning hum in the background, that hum is one of the most consistent things in the whole set — far more consistent than any single word.
So the model learns it, and it comes back on every conversion, in a voice that is otherwise correct. That is a maddening bug to chase later, because the model is working exactly as designed.
| Where the noise is | What it costs you |
|---|---|
| In the source vocal you convert | One conversion is affected |
| In the training dataset | Every conversion that model ever makes |
What to remove, and what to leave
Worth removing: steady hiss, electrical hum, air conditioning, fan noise, computer whirr, traffic rumble. These sit there unchanging, they tell you nothing about the voice, and they are exactly what a noise remover is built for.
Not a noise problem: other people talking, background music, and room echo. Those need different tools — the separator for voices and music, the reverb remover for echo.
You can absolutely overdo this. Push a noise remover too hard and it leaves a metallic, underwater sort of sound behind — and the model learns that just as faithfully as it would have learned the hiss. A slightly noisy natural recording usually trains better than a heavily scrubbed one. If the cleaned file sounds worse to you, it will train worse too.
A sensible order
- Cut what does not belong — other speakers, music, long silences, clipped passages.
- Separate if the voice is over music, so you are left with the voice alone.
- Take the echo off if the recording has an obvious room sound.
- Denoise last, gently, and only if there is a steady noise floor left to remove.
- Listen end to end. Anything that makes you wince is something the model is about to learn.
The dataset preparation guide covers how much audio to aim for; this page is about the quality of those minutes.
Common questions
Does background noise really affect RVC training?
Yes. Steady noise is one of the most consistent things in a dataset, which makes it one of the easiest things for a model to learn.
Before or after training?
Before, always. Afterwards the noise is entangled with the voice and cannot be cleanly separated.
How clean is clean enough?
If you would be happy listening to the dataset end to end, it is clean enough. Chasing silence past that point usually costs more in artifacts than it gains.
Ten free minutes on the noise remover, with its own separate allowance.
Open the noise remover