Train an RVC voice model

Training learns a target voice from a prepared collection of audio. Use this guide when you want to build a model from your own or authorized recordings. If you already have an inference model, see RVC models and files to start using it.

Plan your first run

Prepare the audio with the dataset guide, then select it in Training. Give the experiment a unique name and choose one short recording outside the training set for listening tests. Start with a modest run that can complete all stages and produce a usable export. This gives you a baseline for changes to data or settings.

Record the application version, dataset revision, model format and sample rate, embedder, pitch method, pretrained pair, batch size, and planned save interval. Add the name of the held-out test clip. If a later run sounds different, these notes help you identify what changed.

Choose compatible settings

Setting	What to check
Model format, vocoder, and sample rate	Select a combination supported by the app and intended playback client. A fork-specific architecture can limit where your model loads.
Pitch guidance and extraction	Pitch guidance is useful for modelling pitch-dependent material, especially singing. Pick an extractor available for training in your build.
Embedder	This defines the representation extracted from audio. Record the choice so you can use the matching representation in supported inference clients.
Pretrained generator and discriminator	Use a matched pair compatible with the selected architecture, rate, and pitch configuration. Generator and discriminator files serve different roles in training.
Batch size	Controls how much data is processed per update. Begin with a setting your GPU can fit and lower it if training runs out of memory.
Target and checkpoint interval	An epoch is approximately one pass through the prepared dataset. Steps are individual training updates. Save periodically so you can compare exports and recover after an interruption.

Use the pretraining choices shown as compatible by your installed build. In Ilaria RVC 2.1.0, for example, the project documents RefineGAN training at 32 kHz with RVC V2 and pitch guidance. The same release also documents a downloader for compatible pretrained pairs. Ilaria RVC project documentation, Applio training guide

Run the stages in order

Stage	What it produces	What to check
Dataset preprocessing	Audio prepared for the experiment's model configuration.	The log names the intended folder and reports successful processing for usable files. Check skipped files before continuing.
Feature extraction	Content features, plus pitch data when required by the chosen configuration.	Extraction finishes for the current experiment using the selected embedder and pitch method.
Model training	Checkpoints and inference exports according to the configured save options.	Training steps advance and expected files are saved. Read the first error if progress stops.
Retrieval index	A searchable collection of dataset features for optional use during conversion.	The index belongs to this experiment and matches the exported voice. If you skipped it, test the model independently.

Use the automatic workflow where available, or finish and verify each stage before starting the next. A changed dataset or sample rate can invalidate prepared audio and extracted features. A changed embedder requires new feature extraction. When testing an incompatible configuration, start a separate experiment so existing caches and checkpoints stay available. RVC Project training notes, RVC Project CLI stages

Follow progress in Ilaria RVC

• Use Training Dashboard to check step count, time per epoch, and loss curves. A curve can help explain a stalled run, while listening tests tell you how the output sounds.
• Configure training previews with a constant test clip so the app can generate comparable conversions as checkpoints are saved.
• If you use the Training Queue, set a distinct experiment name and dataset for each job. Before resuming or retrying, confirm which stages the job completed and where it will save output.
• When your release exposes Stop and save, use it when you need to interrupt and continue later. Wait for the checkpoint to finish writing before closing the app.

Ilaria RVC 2.1.0 supports epoch and total-step targets, including resuming partway through an epoch for step targets. Other frontends can expose different controls and resume behavior. Ilaria RVC project documentation, Applio TensorBoard guide

Compare exports by listening

Open inference with an exported .pth. Use the same short held-out recordings and the same conversion settings to compare models from different checkpoints. If you make an index comparison, hold the checkpoint steady and compare the same clip with retrieval disabled and enabled.

Compare these four qualities across exports.

1. Timbre. Does the converted recording resemble the intended vocal character across the whole phrase?
2. Intelligibility. Are consonants and word endings still easy to understand?
3. Pitch stability. Do notes or syllables jump, warble, or lose their intended inflection?
4. Artifacts. Does the output add buzzing, embedded room noise, harsh tails, or odd transitions?

Keep the best sounding export and its test outputs. More epochs can be worth trying when results improve across several examples. If later exports only improve one clip or begin reproducing flaws from the dataset, return to the best baseline and review the recordings. RVC Project FAQ

Export, resume, or share

• Use in inference. Select the exported model .pth. Keep its corresponding .index alongside it when you want retrieval. Record any required embedder, supported client, and sample-rate information.
• Resume training. Back up the experiment and keep the matching generator and discriminator training checkpoints, commonly named G_*.pth and D_*.pth. Resume with the same experiment and compatible configuration, then inspect the log for the checkpoint and step or epoch you expected. An inference export alone lacks the full training state.
• Share a model. Share only files you have permission to distribute. Include listening examples, compatibility notes, the matching index if useful, and any usage terms the recipient needs.

In Ilaria RVC, recent releases use the model library for inference exports and associated indexes. Export paths and naming can differ across frontends. Save the experiment logs separately if you want to recreate the model later. Ilaria RVC model management, Applio export and resume guide

Diagnose a failed run

Symptom	Check first
Preprocessing found no audio	Verify the selected dataset folder and play an input clip. Check whether your build reads nested folders.
Extraction fails or reports a missing model	Check the selected embedder and pitch model assets, and read the first message in the extraction log.
GPU memory runs out	Reduce batch size, stop competing GPU jobs, and disable GPU dataset caching if your build offers that control. Retry the failing stage.
Training restarts at the beginning	Check that the experiment points to compatible training checkpoints and that any fresh-start option is off.
A .pth will not load for conversion	Check whether you selected a training checkpoint instead of an inference export, and verify architecture and embedder compatibility with the client.
A usable model has no index	Check whether index generation was skipped or failed. Export and test the checkpoint independently, then build its index from the matching experiment if needed.

Keep the first failing message and the configuration associated with the run. For more checks, see Troubleshooting. RVC Project FAQ