<![CDATA[The development of voice based biometric security systems has increased the demand for authentication methods capable of operating accurately and securely in open set speaker verification scenarios. In this scenario, the system is required not only to recognize registered users but also to reject unknown users who are not included in the system database. This study focuses on hyperparameter optimization in a Convolutional Neural Network Embedding based speaker verification system using Mel Frequency Cepstral Coefficient (MFCC) features and speaker embeddings. The optimization process was conducted through several experimental stages, including MFCC parameter tuning, CNN architecture tuning, embedding dimension tuning, and audio augmentation analysis. The dataset consisted of Indonesian speech recordings from 8 registered speakers and 1 unknown speaker, sampled at 16 kHz under controlled recording conditions. The dataset was divided into training, enrollment, and testing subsets to support open set speaker verification evaluation and reduce data leakage. System performance was evaluated using accuracy, validation loss, False Acceptance Rate (FAR), False Rejection Rate (FRR), best threshold, and inference time. The experimental results show that the best configuration was achieved using the MFCC-C parameters (N_MFCC = 40, N_FFT = 1024, HOP_LENGTH = 256, N_MELS = 40), the CNN-E architecture with three convolution blocks (32-64-128), an embedding dimension of 64, and lightweight augmentation consisting of noise injection, pitch shifting, and time stretching. This configuration achieved stable system performance with a test accuracy of 96.43% and a FAR of 8.7%, while maintaining lightweight computational complexity and real time inference capability. The results also indicate that excessive augmentation may increase embedding overlap between speakers, thereby reducing system security performance. However, the study was conducted on a limited scale dataset and has not yet evaluated robustness against spoofing attacks, replay attacks, or adversarial synthesized voice attacks. Overall, the study indicates that hyperparameter optimization influences the balance between accuracy, computational efficiency, and biometric security performance in lightweight CNN based voice biometric authentication systems under limited scale evaluation conditions.]]>
