marytts.unitselection.select.viterbi

Class Viterbi

java.lang.Object

marytts.unitselection.select.viterbi.Viterbi

public class Viterbi
extends Object

Provides support for the Viterbi Algorithm. Implementation Notes

For each candidate for the current unit, calculate the cost between it and the first candidate in the next unit. Save only the path that has the least cost. By default, if two candidates come from units that are adjacent in the database, the cost is 0 (i.e., they were spoken together, so they are a perfect match).

Repeat the previous process for each candidate in the next unit, creating a list of least cost paths between the candidates between the current unit and the unit following it.

Toss out all candidates in the current unit that are not included in a path.

Move to the next unit and repeat the process.

Field Summary

Fields

Modifier and Type	Field and Description
protected int	beamSize
protected double	cumulJoinCosts
protected double	cumulTargetCosts
protected ViterbiPoint	firstPoint
protected JoinCostFunction	joinCostFunction
protected ViterbiPoint	lastPoint
protected org.apache.log4j.Logger	logger
protected int	nJoinCosts
protected int	nTargetCosts
protected StatisticalCostFunction	sCostFunction
protected TargetCostFunction	targetCostFunction
protected float	wJoinCosts
protected float	wSCosts
protected float	wTargetCosts

Constructor Summary

Constructors

Constructor and Description
Viterbi(List<Target> targets, UnitDatabase database, float wTargetCosts, float wSCosts, int beamSize) Creates a Viterbi class to process the given utterance.
Viterbi(List<Target> targets, UnitDatabase database, float wTargetCosts, int beamSize) Creates a Viterbi class to process the given utterance.

Method Summary

Methods

Modifier and Type	Method and Description
void	apply() Carry out a Viterbi search in for a prepared queue of ViterbiPoints.
List<SelectedUnit>	getSelectedUnits() Collect and return the best path, as a List of SelectedUnit objects.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

beamSize

protected int beamSize

wTargetCosts

protected final float wTargetCosts

wJoinCosts

protected final float wJoinCosts

wSCosts

protected final float wSCosts

firstPoint

protected ViterbiPoint firstPoint

lastPoint

protected ViterbiPoint lastPoint

targetCostFunction

protected TargetCostFunction targetCostFunction

joinCostFunction

protected JoinCostFunction joinCostFunction

sCostFunction

protected StatisticalCostFunction sCostFunction

logger

protected org.apache.log4j.Logger logger

cumulJoinCosts

protected double cumulJoinCosts

nJoinCosts

protected int nJoinCosts

cumulTargetCosts

protected double cumulTargetCosts

nTargetCosts

protected int nTargetCosts

Constructor Detail

Viterbi

public Viterbi(List<Target> targets,
       UnitDatabase database,
       float wTargetCosts,
       int beamSize)

Creates a Viterbi class to process the given utterance. A queue of ViterbiPoints corresponding to the Items in the Relation segs is built up.

Parameters:

targets - targets

database - database

wTargetCosts - wTargetCosts

beamSize - beamSize

Viterbi

public Viterbi(List<Target> targets,
       UnitDatabase database,
       float wTargetCosts,
       float wSCosts,
       int beamSize)

Creates a Viterbi class to process the given utterance. A queue of ViterbiPoints corresponding to the Items in the Relation segs is built up.

Parameters:

targets - targets

database - database

wTargetCosts - wTargetCosts

wSCosts - wSCosts

beamSize - beamSize

Method Detail

apply

public void apply()
           throws SynthesisException

Carry out a Viterbi search in for a prepared queue of ViterbiPoints. In a nutshell, each Point represents a target item (a target segment); for each target Point, a number of Candidate units in the voice database are determined; a Path structure is built up, based on local best transitions. Concretely, a Path consists of a (possibly empty) previous Path, a current Candidate, and a Score. This Score is a quality measure of the Path; it is calculated as the sum of the previous Path's score, the Candidate's score, and the Cost of joining the Candidate to the previous Path's Candidate. At each step, only one Path leading to each Candidate is retained, viz. the Path with the best Score. All that is left to do is to call result() to get the best-rated path from among the paths associated with the last Point, and to associate the resulting Candidates with the segment items they will realise.

Throws:

SynthesisException - if for any part of the target chain, no candidates can be found

getSelectedUnits

public List<SelectedUnit> getSelectedUnits()

Collect and return the best path, as a List of SelectedUnit objects. Note: This is a replacement for result().

Returns:

the list of selected units, or null if no path could be found.