Multi-Source Domain Adaptation for Text-Independent Forensic Speaker Recognition

Adapting speaker recognition systems to new environments is a widely-used technique improve well-performing model learned from large-scale data towards task-specific small-scale scenarios. However, previous studies focus on single domain adaptation, which neglects more practical scenario where training are collected multiple acoustic domains needed in forensic Audio analysis for offers unique challenges with multi-domain due location/scenario uncertainty and diversity mismatch between reference naturalistic field recordings. It also difficult directly employ domain-specific train complex neural network architectures performance loss. Fine-tuning commonly-used method adaptation order retrain the weights initialized well-trained model. Alternatively, this study, three novel methods based adversarial training, discrepancy minimization, moment-matching approaches proposed further promote across domains. A comprehensive set of experiments conducted demonstrate that: 1) diverse do impact performance, could advance research audio forensics, 2) learns discriminative features invariant shifts domains, 3) discrepancy-minimizing achieves effective simultaneously 4) along dynamic distribution alignment significantly promotes each domain, especially LENA-field noise compared all other systems. Advancements shown here therefore helper ensure consistent operational forensics.