Prepare a voice dataset

Every clip teaches the model something about its target voice. Before training, decide which parts of that voice you want the model to handle, then keep recordings that demonstrate those parts clearly. Start with material you have permission to use for training, and keep a separate copy of the original files.

Choose recordings for your goal

Goal	What to collect
Conversational speech	Questions, statements, longer sentences, and short responses with varied consonants and vowels.
Expressive speech	The intended delivery styles at both quiet and emphatic levels. Record variations across several phrases.
Singing	Sustained vowels, note changes, and examples near the comfortable high and low ends of the intended range.
More than one language	Clear examples of the sounds and phrasing that differ across languages, recorded by the same speaker.

A clean, varied set of recordings is a better starting point than a larger collection of repeated phrases or damaged audio. There is no duration that guarantees a good result. Audit what the recordings cover before deciding whether you need more material. Applio dataset guide

Record or select clean material

Use one speaker for a first model. Exclude overlapping voices, crowd responses, and dialogue where the speaker changes mid-clip.
Listen for fan noise, echoes, music, and digital distortion. A clip that sounds acceptable in a loud mix can still contain problems that are obvious when the voice is isolated.
If you are recording a new session, use a consistent mic position and leave peak headroom. Test the loudest phrase before recording the full session.
When combining sessions, listen to them side by side. Keep useful variation in the voice and discard abrupt differences caused by heavy effects, room sound, or an unreliable microphone.
Remove duplicate takes if they add little variety. Put different sessions or passages aside for evaluation before editing adjacent cuts into the training folder.

Clean a working copy

Isolate the target voice. If the source includes an accompaniment, try separation and audition the resulting vocal stem. Reject sections with instrument bleed, backing vocals, or obvious separation artifacts.
Make cuts at natural boundaries. Trim long dead air and remove unusable sections. Listen across each cut for lost consonants, abrupt changes, or a second speaker entering.
Apply only the cleanup the recording needs. Compare any denoising pass against the original. If the cleaned version swallows consonants, blurs syllables, or introduces a pulsing sound, reduce the processing or replace the take.
Check levels across files. Keep enough headroom for loud passages and balance major recording-level differences. Lowering the volume of an already clipped file leaves the damaged peaks in the recording.
Export in a lossless format your build can read. WAV is a practical working format. Re-encoding a compressed file as WAV preserves the result of that decode and cannot restore lost detail.

The training engine may divide files into its own segments and resample them to the selected model configuration. Keep usable words and phrases intact during manual editing and let the training preprocessing stage handle its own segmentation. AI Hub guide to isolation and cleanup, RVC Project training notes

Organize the dataset and listening tests

Keep comparison inputs, discarded recordings, and converted output outside the training folder. This example uses separate folders for originals, training, and evaluation.

my-voice-project/
  originals/
  training-audio/
    session-a-01.wav
    session-a-02.wav
    session-b-01.wav
  evaluation/
    speech-test.wav
    expressive-test.wav
  notes.md

Point preprocessing at training-audio/. Keep evaluation/ separate. A useful comparison set might include an unfamiliar spoken phrase with sharp consonants, a quiet passage, and a short expressive phrase. For a singing model, include a held note and a phrase crossing registers. Hold back whole passages or sessions when you can, so the test audio does not duplicate nearby training clips.

If you train with an interface that supports more than one speaker, follow its speaker-folder and ID rules. A single-speaker folder with audio files directly inside it is easier to audit for a first run. Applio dataset guide

Review the dataset in Ilaria RVC

Recent Ilaria RVC builds let you drag audio or a dataset ZIP into the training interface and manage imported datasets there. Select the intended dataset and run the Dataset Checker. Use its short previews to inspect files flagged for silence, clipping, inconsistent levels, or decoding problems. The Audio Analyzer can help compare suspect files using waveforms, pitch, and level measurements. Confirm the diagnosis by listening before removing otherwise useful material. Availability and control labels follow your installed release. Ilaria RVC project documentation

What you notice	What to do before training
File will not decode	Check that it plays in an editor. Re-export from the original if possible and verify the copy opens.
Sections of near-silence	Audition the flagged range and trim empty spans. Keep natural pauses that help preserve phrase boundaries.
Distorted or flattened peaks	Compare with the original recording. Replace an audibly clipped take when you have a clean alternative.
Background sounds follow the vocal	Listen to the isolated stem in headphones and exclude the affected parts when separation sounds unstable.
One session is much quieter	Check whether the speaker's performance is intentionally quiet. Adjust recording gain differences while retaining useful natural dynamics.

Before continuing, play a few clips from each session, confirm you selected the prepared folder, and record the folder version in your notes. Then run a training experiment.