Speaker identification remains critical in biometric authentication systems, requiring robust feature extraction strategies that capture speaker-specific vocal characteristics. This study introduces a hybrid deep learning architecture integrating Convolutional Neural Networks (CNNs) with Reinforcement Learning (RL) for confidence-aware speaker identification. Two feature extraction methodologies were compared: Method 1 employed Mel-spectrogram representations (80 bins, 20-8000 Hz) with self-attention mechanisms, while Method 2 utilized Continuous Wavelet Transform with Morlet wavelets (128 scales). Both methods were implemented as hybrid CNN-RL architectures and compared against CNN-only baselines. Frameworks were evaluated on LibriSpeech dev-clean dataset (2,703 audio files, 40 speakers) through stratified 5-fold cross-validation. ANOVA assessed discriminative capacity of 22 acoustic features. ANOVA revealed 21 of 22 features demonstrated significant discriminative power (p < 0.05), with entropy exhibiting the strongest effect (F = 39.79, η² = 0.37). Method 1 achieved 87.60% accuracy (95% CI: 83.60%, 91.95%) and ROC-AUC of 99.54%. Method 2 attained 77.60% accuracy (95% CI: 73.12%, 82.08%) with ROC-AUC of 98.21%. Ablation studies demonstrated that RL integration provided statistically significant improvements over CNN-only baselines (Method 1: +2.80%, p = 0.0142; Method 2: +3.40%, p = 0.0089) with reduced performance variability. Both architectures demonstrated balanced metrics with Matthews Correlation Coefficients exceeding 0.76. Q-learning integration enabled adaptive decision-making for classification uncertainty, with ablation studies confirming quantitative contributions over conventional CNNs. Results demonstrate that Mel-scale representations with attention mechanisms provide superior discriminative capacity compared to fixed-resolution wavelet transforms for robust biometric authentication.
Heir et al. (Thu,) studied this question.