private float ConvertScore(float val)
{
// Convert to the appropriate base.
val = LogMath.LnToLog(val);
// TODO: Need to use mean and variance transforms here
if (float.IsNaN(val))
{
this.LogInfo("gs is Nan, converting to 0");
val = LogMath.LogZero;
}
if (val < DistFloor)
{
val = DistFloor;
}
return(val);
}
/**
* /// Calculate the score for this mixture against the given feature. We model the output
* /// distributions using a mixture of Gaussians, therefore the current implementation is simply
* /// the computation of a multi-dimensional Gaussian. <p/> <p><b>Normal(x) = exp{-0.5/// (x-m)' *
* /// inv(Var)/// (x-m)} / {sqrt((2/// PI) ^ N)/// det(Var))}</b></p>
* /// <p/>
* /// where <b>x</b> and <b>m</b> are the incoming cepstra and mean vector respectively,
* /// <b>Var</b> is the Covariance matrix, <b>det()</b> is the determinant of a matrix,
* /// <b>inv()</b> is its inverse, <b>exp</b> is the exponential operator, <b>x'</b> is the
* /// transposed vector of <b>x</b> and <b>N</b> is the dimension of the vectors <b>x</b> and
* /// <b>m</b>.
*
* /// @param feature the feature to score
* /// @return the score, in log, for the given feature
*/
public float GetScore(float[] feature)
{
// float logVal = 0.0f;
var logDval = LogPreComputedGaussianFactor;
// First, compute the argument of the exponential function in
// the definition of the Gaussian, then convert it to the
// appropriate base. If the log base is <code>Math.E</code>,
// then no operation is necessary.
for (var i = 0; i < feature.Length; i++)
{
var logDiff = feature[i] - MeanTransformed[i];
logDval += logDiff * logDiff * PrecisionTransformed[i];
}
// logDval = -logVal / 2;
// At this point, we have the ln() of what we need, that is,
// the argument of the exponential in the javadoc comment.
// Convert to the appropriate base.
logDval = LogMath.LnToLog(logDval);
// System.out.println("MC: getscore " + logDval);
// TODO: Need to use mean and variance transforms here
if (float.IsNaN(logDval))
{
this.LogInfo("gs is Nan, converting to 0");
logDval = LogMath.LogZero;
}
if (logDval < DistFloor)
{
logDval = DistFloor;
}
return(logDval);
}
/**
* Converts the probability from -ln to logmath
*
* @param lnProb the probability to convert. Probabilities in the arpa format in negative natural log format. We
* convert them to logmath.
* @return the converted probability in logMath log base
*/
private float ConvertProbability(float lnProb)
{
return(_logMath.LnToLog(-lnProb));
}
