Outlier elimantion algorithms for voice conversion.
The aim of outlier elimination is to detect source and target pairs that might result in reduced conversion performance or artefacts.
The best working method so far has been the GaussianOutlierEliminator.
This method fits a single Gaussian to the difference distributions of various source and target acoustic features (LSFs, duration, f0, energy).
Then, the pairs with significant difference as compared to distribution mean are eliminated.
The voice conversion training step can also be set to eliminate too close pairs, forcing an average amount of unlikeliness in the training data.