Speaker discrimination using formant trajectories from casework recordings: Can LDA do it?

Formant trajectories have been shown to convey a great deal of speaker-specific information and their speaker-discriminatory potential has been quantified using Linear Discriminant Analysis on laboratory material [16]. This study tests the applicability of LDA on three sets of real-case forensic recordings. Given the limitations of LDA, we used the actual formant trajectory values (F1–F3) and coefficients of the quadratic and cubic fit. As for classification rate, our results indicate that LDA performs comparably to the studio condition, with quadratic fit being the most convenient way of parametrizing the trajectory. However, LDA performed well above chance when discriminating between recordings of the same speaker; it is especially this inability to “identify” the same speaker which makes the use of LDA in forensic practice not recommendable.