Glossary

Term	Meaning
Inference	Using a trained model to convert a recording.
Checkpoint	File containing learned model weights. Resume checkpoints and inference exports have different roles.
Embedder	Component that extracts a representation of the source audio for the conversion engine.
Index	Optional collection of learned feature vectors used during retrieval.
Index ratio	How strongly an implementation uses retrieved features during conversion.
F0	Fundamental frequency, used to represent pitch over time.
Pitch extractor	Algorithm that estimates F0 from a recording.
Semitone	One twelfth of an octave. A change of 12 semitones is one octave.
Vocoder	The part of the pipeline that generates the waveform from its conditioning inputs.
Sample rate	The number of audio samples per second. It is distinct from model version and pitch transposition.
Pretrain	Starting weights used when training a new model.
Epoch	One pass through the training dataset in the usual training setup.
Step	A training update. Batch size and dataset size affect the relationship between steps and epochs.
Stem	One isolated part of a mix, typically a vocal or instrumental.
Latency	The time between an input event and its audible output.