Effectiveness of Phoneme-Based Speaker Recognition Using Ultra-Short Speech Segments: The Case Study of the Kazakh Language

Abstract

This paper presents an extended study of a phoneme-based approach to speaker identification in the Kazakh language using ultra-short utterances. Unlike the previous study conducted on a corpus of 20 speakers, the present experiment was conducted on 145 speakers, which made it possible to verify the reproducibility of the identified patterns under conditions close to real biometric authentication systems. The material consisted of short words with a Consonant–Vowel–Consonant (CVC) structure. Four models based on the ECAPA-TDNN architecture were developed: three phoneme-based models (for the vowels "A," "O," and "E") and one generalized model (AEO). The data were normalized in terms of loudness and cleaned of leading and trailing pauses using VAD. The training and test corpora did not overlap in terms of speakers. For test utterances ranging from 3 to 24 words, all models exhibited an increase in accuracy following a power-law trend as the utterance length increased. For short utterances, phoneme-based models retained an advantage, particularly the model based on the vowel "E," which is associated with phonetic stability and consistent formant structure. Expanding the corpus to 145 speakers reduced the variability of the results and confirmed the robustness of the observed effects. The practical value of this study lies in the fact that using words dominated by the vowel "E" in short voice passwords can improve identification accuracy without increasing command length or the amount of training data, which is especially important for voice authentication systems in low-resource languages.