Formant contours in Czech vowels: speaker-discriminating potential

The usefulness of dynamic formant properties for speaker discrimination was demonstrated on English, mostly exploiting long vowels or diphthongs, characterized both by actual measures along the formant tracks and coefficients from polynomial regression on these tracks. This study applies this paradigm to Czech, using less tightly controlled (more forensically realistic) material and taking into account the specific properties of the Czech vocalic system in which long vowels are much rarer than short ones. When all vowels are pooled together, the best results are achieved for unstressed vowels in asymmetrical CVC contexts. When individual vowels are considered separately, classification rates are best for long [i:] and [a:], but, most importantly, short vowels also show promising results. The performance of actual formant values and regression coefficients as predictors in discriminant analysis appears comparable.