A hybrid CNN and reinforcement learning framework for speaker identification using Mel-Spectrogram and continuous...

Area of Science:

• Biometrics and Security
• Artificial Intelligence and Machine Learning
• Signal Processing

Background:

• Speaker identification is crucial for biometric authentication.
• Robust feature extraction is needed to capture unique vocal characteristics.
• Current methods require advanced deep learning architectures.

Purpose of the Study:

• To develop a novel hybrid deep learning architecture for confidence-aware speaker identification.
• To compare Mel-spectrograms with self-attention against Continuous Wavelet Transform for feature extraction.
• To evaluate the impact of Reinforcement Learning (RL) integration on Convolutional Neural Networks (CNNs).

Main Methods:

• A hybrid CNN-RL architecture was developed for speaker identification.
• Two feature extraction methods were employed: Mel-spectrograms (Method 1) and Continuous Wavelet Transform (Method 2).
• The LibriSpeech dataset was used with 5-fold cross-validation, and ANOVA assessed feature discriminative power.

Main Results:

• Mel-spectrograms with attention (Method 1) achieved 87.60% accuracy and 99.54% ROC-AUC.
• Continuous Wavelet Transform (Method 2) achieved 77.60% accuracy and 98.21% ROC-AUC.
• RL integration significantly improved CNN-only baselines (p < 0.05) and reduced performance variability.

Conclusions:

• Mel-scale representations with attention mechanisms provide superior discriminative capacity for speaker identification.
• Hybrid CNN-RL architectures enhance robustness and reduce uncertainty in biometric authentication.
• The proposed methods offer a significant advancement in confidence-aware speaker identification systems.