Semi-automatic approach utilizing Siamese Neural Network for forensic voice comparison

Forensic Voice Comparison (FVC) remains a critical yet challenging task in digital forensics, often hindered by manual subjectivity, background noise, and speaker variability. This paper presents a novel semi-automatic FVC framework based on Siamese Neural Networks (SNN), a discriminative metric-learning architecture combined with stationary noise reduction for robust voice similarity assessment. Proposed framework leverages the SNN’s ability to learn a shared embedding space where Euclidean distance reflects speaker identity. Using a jurisdiction-specific dataset of 3899 Australian English speech samples (FLAC format), proposed framework achieves 96.02% accuracy, 94.00% precision, and 92.10% recall in distinguishing same vs different speaker pairs. The proposed framework is evaluated against strong baselines including Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BiLSTM), Gaussian Mixture Model-Universal Background Model (GMM-UBM), and validated via 5-fold cross-validation (mean ± std. dev.) to ensure statistical robustness. Proposed framework fills a critical gap in forensic phonetics by demonstrating that lightweight, interpretable, pairwise deep learning models can outperform complex generative or ensemble systems in real-world FVC scenarios. All preprocessing, training protocols, and hyperparameters are documented for reproducibility.