<![CDATA[Speaker gender classification is a critical component in automatic speech processing systems, particularly for enabling personalized voice-based services. This study proposes a speaker gender classification system for Indonesian audio recordings by combining Mel-Frequency Cepstral Coefficients (MFCC) feature extraction with a Bidirectional Long Short-Term Memory (BiLSTM) deep learning architecture optimized using the Nadam optimizer. The dataset utilized is Mozilla Common Voice version 24.0 Indonesian, comprising 21,545 gender-labeled audio clips with a distribution of 13,965 male clips (64.8%) and 7,580 female clips (35.2%). Audio preprocessing includes resampling to 16,000 Hz, Z-score normalization, and conversion from stereo to mono. Feature extraction yields 120 features per time step, composed of 40 MFCC coefficients, 40 Delta, and 40 Delta-Delta components, producing an input matrix of 501×120 per audio clip. Class imbalance is addressed through the class weight technique. The dataset is partitioned using a stratified 70:15:15 split for training, validation, and testing. The proposed BiLSTM model architecture consists of two bidirectional LSTM layers (128 and 64 units), a fully connected dense layer, and a sigmoid output layer, supported by Dropout and Batch Normalization regularization techniques. Evaluation results demonstrate that the model achieves an accuracy of 98.58%, F1-Score of 98.58%, sensitivity of 98.90%, and specificity of 97.98%, surpassing all prior Indonesian-language gender classification studies and demonstrating that the integration of delta-features and bidirectional architectures significantly overcomes previous limitations in regional accent variability, thereby providing a robust foundation for future Indonesian speech-based intelligent systems.]]>
