Speaker Identification Using Voice Quality Features: A Psychoacoustic and Machine Learning Approach

Area of Science:

• Phonetics and Speech Science
• Acoustic Analysis
• Biometrics

Background:

• Speaker identification relies on unique voice characteristics.
• Psychoacoustic models offer a framework for analyzing voice quality.
• Understanding sex-specific voice features is crucial for accurate speaker recognition.

Purpose of the Study:

• To assess the effectiveness of psychoacoustic voice features for closed-set speaker identification.
• To determine the discriminative power of static and dynamic voice features.
• To identify key features for distinguishing individual Persian speakers.

Main Methods:

• Collected speech recordings from 120 native Persian speakers (60 male, 60 female).
• Extracted 26 psychoacoustic voice features (13 static, 13 dynamic using coefficient of variation).
• Utilized a Random Forest classifier for speaker separability analysis.

Main Results:

• Combined features achieved 90.15% accuracy for males and 90.74% for females.
• Static features were more effective for females (91.24%) than dynamic features (76.52%).
• Dynamic features improved accuracy for males (90.15%) compared to static alone (89.53%).
• Key features differed by sex: males (H1*-H2*, CoVH1*-H2*, f0), females (f0, CoVf0, F3).

Conclusions:

• Psychoacoustic voice features are effective for Persian speaker identification.
• The utility of dynamic voice features for speaker discrimination is sex-dependent.
• Temporal variability enhances male speaker recognition, while stable features are more critical for females.
• Findings inform forensic phonetics and speaker recognition system development.