Single and batch inference

Establish a reference

Open Single Inference, choose a voice and load a short, dry clip. Set Pitch and Index Ratio to zero. Use the embedder expected by the model, start with RMVPE, and leave optional effects off. Save a WAV and compare it with your input at matching loudness.

Tune in separate passes

If a matching index is available, raise retrieval in small increments. Listen for clearer voice character and possible muffling. Lower Protect to give consonants and breaths more protection in pitch-guided conversion. The effect depends on retrieval, which contributes only above zero. Compare pitch settings separately.

Process multiple files

In Batch Inference, group recordings by model and pitch setting. Use a separate destination folder. Run one representative clip first, then check the batch log and listen to a result from each group.