An epoch is one complete pass through the prepared dataset. More passes increase training time, but they do not guarantee a better voice. The right number depends on the recording length, quality, variation, and how quickly you need a usable result.
NiceVois's starting recommendations
| Usable audio duration | Starting recommendation | What to expect |
|---|---|---|
| Under 1 minute | 25 epochs | A fast experiment; the source may be too short for a reliable model. |
| 1 to under 5 minutes | 100 epochs | The standard public-beta starting point. |
| 5 to under 10 minutes | 150 epochs | A practical starting point; longer clean audio supplies more source variety. |
| 10 minutes or more | 200 epochs | A balance-funded starting point, subject to the live feasibility check. |
Every job receives an exact quote before training. Balance-funded runs may select up to 500 epochs, and the uploader enforces the audio and feasibility limits shown for your workspace.
Undertraining versus overtraining
With too few epochs, the model may not capture a stable voice identity. With excessive training, it can overfit the particular recording and reproduce its quirks instead of generalizing cleanly. Source quality remains more important than chasing a large number.
Use checkpoints when time matters
The training workspace can publish an inference-ready checkpoint when one becomes available. That gives you a model to test before the final epoch target finishes. A checkpoint is useful for urgent work and for deciding whether the source audio is promising.
A practical decision rule
- Start with the duration-based recommendation.
- Test the resulting model on speech or singing that was not in the dataset.
- If identity is weak but the source is clean, consider a longer run.
- If artifacts mirror the dataset, improve the audio before adding more epochs.
Simple answer: start around 100 epochs below five minutes, 150 below ten minutes, and 200 for ten minutes or more. Go higher only when the result sounds undertrained. Use 25 only when you need a quick short-audio experiment.
Add your audio and the epoch dial will show the recommendation for its duration.
Train an RVC model from audio