Enhancing the robustness of brain-inspired perception decision machine for fake speech detection

Introduction The rapid development of speech synthesis, voice conversion, and large-scale generative models has made fake speech more natural and difficult to detect. Existing detectors can respond differently to the same sample, yet fixed fusion does not use the state of each detector to decide how much that detector should contribute. Methods Building on the brain-inspired perception-decision framework of Feng et al., we propose the Dynamic Routing Perception-Decision Machine (DR-PDM). The method retains detector embeddings, class probabilities, one-class Softmax scores, and evidential uncertainty as detector-state information, and uses these states to generate sample-dependent fusion weights. For internally evaluated systems, operating thresholds are selected by a reproducible discrete equal-error-rate rule on development scores and frozen before evaluation. WaveFake uses its own nonoverlapping training, development, and evaluation splits: all four experts and the Router/fusion backend are trained on WaveFake training data, and no ASVspoof-trained expert weights are used. It is therefore a within-dataset experiment. Results At the applied thresholds, DR-PDM obtains precision/missing-rate pairs of 99.995%/2.085%, 99.815%/0.930%, and 100.000%/0.007% on ASVspoof2019 LA, ASVspoof2021 LA, and WaveFake, respectively. The zero-false-positive WaveFake result has a two-sided 95% exact precision interval of 99.986% –100.000%. A serial FP32 benchmark of the complete 1.292-billion-parameter system on an NVIDIA RTX 4090 measures 102.919 ms mean forward-pass latency for a 4.0375-s input, or 4.260 times the average single-expert latency. Discussion The evidence supports detector-state routing under the evaluated protocols, but does not establish overall error reduction, cross-dataset WaveFake transfer, or low-cost deployment.