This study examines the effectiveness of long-term formant frequency distributions (LTFDs), Mel-frequency cepstral coefficients (MFCCs), and their combined application in distinguishing Kurdish-Persian bilingual speakers. Speech samples were collected from 20 early male bilingual speakers who read the fable ' The North Wind and the Sun ' in Kurdish (Sorani dialect) and Persian. The Random Forest algorithm was employed to analyze the data. Feature importance for formant frequencies and MFCCs was evaluated using the mean decrease in accuracy metric. The results indicated that LTFD measures provided moderate accuracy in speaker differentiation, reflecting their capacity to capture vowel-related articulatory patterns. In contrast, MFCCs demonstrated superior performance, effectively encoding spectral and speaker-specific characteristics. When LTFDs and MFCCs were combined, system accuracy was slightly improved compared to using MFCCs alone. This marginal enhancement underscores the potential benefits of integrating LTFDs with MFCCs in forensic voice comparison, where even small gains can have significant practical implications. The findings contribute to a deeper understanding of bilingual speaker variability and provide insights for optimizing speaker identification systems in bilingual contexts.
