Lossless training source

WAV to PTH: train an RVC model from a WAV file

When you have both an original WAV and a compressed copy of the same recording, the WAV is normally the better training source because it avoids an additional lossy encoding stage.

What makes a WAV useful for RVC?

The container alone does not guarantee good training data. A clean WAV with one voice, consistent volume, and little echo is useful. A WAV containing a full music mix, clipping, several speakers, or aggressive noise-reduction artifacts is still a weak dataset.

• Use the original recording or clean isolated vocal whenever possible.
• Remove silence only when it dominates the file; natural pauses are not inherently harmful.
• Remove other speakers, backing vocals, and instrumental bleed.
• Keep natural variation in pitch, words, pace, and expression.

Do you need to resample or split the WAV yourself?

No. The cloud pipeline reads the uploaded source, slices it into workable segments, normalizes and resamples the training material as required, extracts RMVPE pitch and HuBERT features, and then trains the pinned RVC v2 model. Manual splitting is unnecessary for this interface.

Verified WAV workflow

The project’s real end-to-end proof used a 33,189,998-byte WAV. It passed browser upload, private dataset preparation, Kaggle GPU submission, native RVC v2 training, model validation, and named .pth download. A later 100-epoch run produced 53 prepared segments and a 57,590,877-byte inference model that loaded in the native RVC synthesizer.

These numbers verify compatibility and artifact production; they are not a promise that every recording will produce the same perceptual quality or runtime. See the testing methodology.

WAV upload checklist