ape.plot: Function to make a Applied Probability of Error plot
In davidavdav/sretools.R: Compute error metrics and make DET plots for NIST-like Speaker Recognition Evaluations

Description Usage Arguments Details Value Author(s) References See Also Examples

These function plots one or more APE plots, given a list of det objects, or a single NBE plot. In an APE plot one can observe the loss of information as a function of effective prior in a detection experiment. Both the detection and calibration information loss are shown.

In an NBE plot, the error rates are normalized such that a non-informative trivial detector that outputs LR=1 for all input, has error 1. This way of plotting zooms in on low-prior odds, allowing inspection of the calibration performance at low priors.

1
2
3

ape.plot(data, legend = "Ape plot", bar = T, bw = F)

nbe.plot(data)

`data`	An object (or list) of class `det`
`legend`	An array of strings to be used as legend
`bar`	Do we need to make a bar plot as well
`bw`	Use gray shades instead of colours

The data is either a list of objects or a single object of class det. For each det, an APE plot is drawn. In the APE plot one can observe the total error as a function of the ‘log prior odds’ a.k.a. ‘effective prior odds’—which are determined by the cost function. The red curve shows the probability of error given the supplied scores, interpreted as log-likelihood-ratios. The green curve shows the same values after optimal calibration using isotonic regression to convert (uncalibrated) scores into log-likelihood ratios. (For a fixed operation point cost function such as Cdet, this is the same as computing minCdet.) The black dashed curve indicated the performance of a system based on effective prior alone, i.e., ignoring information in the observed data.

If bar==TRUE, bars indication the area under the curves is plotted beneath the APE plots. This can readily be interpreted as Cllr (red bar) and minCllr (green), measured in bits.

The value returned are the graphical parameters par() used. This may not be useful in any way.

David A. van Leeuwen

Niko Br\"ummer and Johan du Preez, “Application-independent evaluation of speaker detection,” Computer, Speech and Language 20, 230–275, (2006). David van Leeuwen and Niko Br\"ummer, “An Introduction to Application-Independent Evaluation of Speaker Recognition System,” LNCS 4343 (2007).

plot.det

## example SRE 2008 data set for short2-short3 condition
data(tno.2008)
## select NIST condition 6
x <- subset(tno.2008, mtype=="phonecall" & ttype=="phonecall" &
  tmic=="phn")
## select low-error-favourite NIST condition 7
y <- subset(x, mlang=="eng" & tlang=="eng")
## compare the two
ape.plot(list(det(x), det(y)), legend=c("all phn", "english only"))
## look at the NBE plot for the English trials
nbe.plot(det(y))