The present paper suggests a secure speaker verification system that can resist replay attacks, one of the most common threats to biometric voice security. The system combines Perceptual Linear Predictive Cepstral (PLPC) features with a K-means clustering algorithm for differentiating between real and replayed audio samples. PLPC features are extracted, normalized, and partitioned into cluster models of authentic speaker profiles during training. In experimentation, the system pairs test features to these clusters via minimum distance criteria and assigns speaker identity accordingly. Performance was assessed using the AVSpoof database, which included real and spoofed audio captured through smartphones and laptops. Experimental findings over ten speakers (five male, five female) exhibit that the system attains 100% accuracy in classification of valid samples, while keeping drastically lower diagonal values in confusion matrices for replayed inputs, ensuring its robust spoofing resistance. Correlation analysis also ensured that high auto-correlation values define consistent valid speech, while low or negative cross-correlations efficiently indicate replayed audio. The suggested strategy thereby generates a computationally efficient and interpretable countermeasure for real-time speaker verification systems with high reliability but without deep learning complexity. In conclusion, the results confirm that the integration of perceptually motivated features with unsupervised clustering significantly boosts voice-based authentication system security and robustness.