ForenVoice-Secure: Robust and Privacy-Aware Audio Data Mining for Forensic Speaker Identification

Speech is now routine evidence in criminal investigations, but forensic audio rarely matches the clean assumptions of standard speaker recognition. Clips are short, noisy, codec-compressed, and channel-mismatched, and they are increasingly exposed to replay and synthetic speech manipulation. Therefore, the cast criminal voice identification is forensic audio data mining, aiming to extract a stable identity structure from heterogeneous and potentially adversarial evidence, while respecting operational and privacy constraints. In this study, a novel ForenVoice-Secure system is proposed, a unified pipeline that combines robust representation learning, spoof-aware decisioning, and privacy-preserving training. Audio is mapped to log-Mel spectrograms and encoded with a CNN, while an LSTM aggregates temporal identity cues from irregular utterances. Robustness is improved through multi-task learning (identity + spoof), adversarial training, and spectro-temporal consistency checks for replay/deepfake artifacts. Privacy is addressed using federated learning, keeping raw recordings local and sharing only model updates. Experiments on VoxCeleb2, ASVspoof 2021, and a forensic-style speaker comparison corpus achieve statistically significant performance gains, 98.43% mean identification accuracy with strong class-balanced performance (macro F1 = 98.10%, precision = 98.22%, recall = 98.01%) and statistically significant gains over strong baselines across repeated folds (F1: p=8.0×〖10〗^(-4); precision: p=1.1×〖10〗^(-3); recall: p=9.0×〖10〗^(-4)). The model remains lightweight (≈4.3M parameters, ≈1.2 GFLOPs per 3 s), enabling near real-time inference with modest overhead from consistency checks (