Revisiting the speaker discriminatory power of vowel formant frequencies under a likelihood ratio-based paradigm: The case of mismatched speaking styles

Abstract

Differentiating subjects through the comparison of their recorded speech is a common endeavor in speaker characterization. When using an acoustic-based approach, this task typically involves scrutinizing specific acoustic parameters and assessing their discriminatory capacity. This experimental study aimed to evaluate the speaker discriminatory power of vowel formants—resonance peaks in the vocal tract—in two different speaking styles: Dialogue and Interview. Different testing procedures were applied, specifically metrics compatible with the likelihood ratio paradigm. Only high-quality recordings were analyzed in this study. The participants were 20 male Brazilian Portuguese (BP) speakers from the same dialectal area. Two speaker-discriminatory power estimates were examined through Multivariate Kernel Density analysis: Log cost-likelihood ratios (Cllr) and equal error rates (EER). As expected, the discriminatory performance was stronger for style-matched analyses than for mismatched-style analyses. In order of relevance, F3, F4, and F1 performed the best in style-matched comparisons, as suggested by lower Cllr and EER values. F2 performed the worst intra-style in both Dialogue and Interview. The discriminatory power of all individual formants (F1-F4) appeared to be affected in the mismatched condition, demonstrating that discriminatory power is sensitive to style-driven changes in speech production. The combination of higher formants ‘F3 + F4’ outperformed the combination of lower formants ‘F1 + F2’. However, in mismatched-style analyses, the magnitude of improvement in Cllr and EER scores increased as more formants were incorporated into the model. The best discriminatory performance was achieved when most formants were combined. Applying multivariate analysis not only reduced average Cllr and EER scores but also influenced the overall probability distribution, shifting the probability density distribution towards lower Cllr and EER values. In general, front and central vowels were found more speaker discriminatory than back vowels as far as the ‘F1 + F2’ relation was concerned.