The d´ model of signal detection applied to speech segmentation

The statistical measure, d\´, from Signal Detection Theory, [Swe64] has been shown to parametrize the "detectability" of signal over noise in a wide variety of perceptual situations. Its usefulness is extended to the problem of quantifying error rates for segmentation of continuous speech. It has often been impossible to accurately compare different machine techniques for segmentation since errors occur as either missing or extra segment boundaries whose rates are related by internal decision thresholds. The basic d\´ model is shown to accurately (>95% confidence) describe the missing versus extra segment trade-off found in at least one, non-trivial, speech segmentation program, [Gol75].