Package marytts.signalproc.adaptation.outlier

Outlier elimination algorithms for voice conversion.

See: Description

Class Summary

Class	Description
BaselineOutlierEliminatorParams	Baseline class for outlier elimination parameters
GaussianOutlierEliminator	Single Gaussian based outlier elimination.
GaussianOutlierEliminatorParams	Parameters for single Gaussian based outlier elimination
GMMOutlierEliminator	TO DO: GMM based outlier elimination
KMeansMappingEliminator	K-Means clustering and mapping based outlier elimination.
KMeansMappingEliminatorParams	This class implements a K-Means clustering and mapping based outlier elimination procedure: - Step1: Cluster source and target acoustic features either jointly or separately - Step2: For each feature, for each source cluster find the most likely target cluster - Step3: For each feature, for each target cluster find the most likely source cluster - Step4: Determine outlier pairs by checking the total number of source-target pairs assigned to clusters other than the most likely cluster which are sufficiently "distant" from the most likely cluster
TotalStandardDeviations	Class for keeping total standard deviations to be used in automatic thresholding in outlier elimation

Package marytts.signalproc.adaptation.outlier Description

Outlier elimination 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.