Abstract: Linear discriminant analysis (LDA) is a commonly used method to evaluate the speaker discriminatory power of formant parameters in forensic voice comparison (FVC). To investigate the sample size (number of speakers) required to ensure the robustness of the classification rate (CR) results (reliably reflecting the relative value of different formants in FVC), in this study, the dynamic features of the first four formants (F1—F4) of the Chinese diphthong /ei/ of 660 young male speakers are extracted. The raw formant trajectories are normalized into ten equal parts, followed by polynomial fitting for dimensionality reduction. The fitting coefficients are used as predictor variables to conduct the LDA. The CR values of different formant parameters under various sample size conditions are systematically analyzed. The results show that when F4 is not considered, at least 40 speakers are needed to stably reflect the relative value of speaker discriminatory power among the individual formants F1—F3. When the combinations of F1—F3 are analyzed or F4 is considered, the minimum sample size will increase significantly.