Dynamic Routing Perception-Decision Machine Evaluates Fake Speech Across Three Detection Benchmarks

A research team led by scientists from Xinjiang University, Naval Aviation University, and Tsinghua University has evaluated a multi-detector framework called the Dynamic Routing Perception-Decision Machine for detecting machine-generated fake speech. The investigation, published in Frontiers in Signal Processing, assesses how combining heterogeneous detector representations using sample-dependent dynamic fusion influences classification trade-offs across established speech benchmarks. Modern developments in speech synthesis, voice conversion, and generative modeling have generated synthetic speech that closely mimics natural human voices, creating significant verification challenges for telecommunications security and automated voice systems. While existing anti-spoofing systems employ distinct acoustic and learned representations, individual detectors frequently exhibit varying sensitivities across different audio samples, leaving fixed ensemble fusion unable to adapt to individual detector confidence levels.

Building directly upon the perception-decision framework previously established by Feng and colleagues, the researchers designed an architecture that organizes multiple specialized detectors under an explicit routing mechanism. Rather than relying on a single neural network or static voting weights, the Dynamic Routing Perception-Decision Machine preserves internal detector representations across four distinct feature streams and computes continuous fusion weights for every evaluated audio sample. The framework extracts detailed detector-state information from each branch, incorporating high-dimensional embedding vectors, binary class probabilities, one-class Softmax normalized similarity scores, and evidence-derived predictive uncertainty. These explicit state variables provide the routing module with a representation of each detector branch’s internal response before producing final fused classification decisions.

The system incorporates four complementary detector branches adapted from published anti-spoofing architectures to capture diverse acoustic properties across speech signals. A phoneme-related branch utilizes an SLS classifier operating on XLS-R speech representations, while a spectrogram branch models temporal-channel relationships using multi-head self-attention mechanisms. A third branch evaluates time-frequency structure through an XLSR-Mamba backbone, and a fourth branch captures raw time-domain waveform cues through a wav2vec2- and Nes2Net-related design. The researchers note that these underlying detector backbones represent adaptations of existing literature rather than novel architectural inventions of the current study, with the investigation focusing primarily on how the routing module interprets their internal states.

To generate sample-specific weights, the architecture maps the intermediate embedding vectors from all four branches into a shared representation space and pairs them with their corresponding detector states. Each branch computes evidence for bona fide and fake speech classes using an evidence head, parameterized under a Dirichlet distribution where smaller total evidence yields higher evidential uncertainty. Normalized similarity scores are simultaneously computed against bona fide class centers using a stabilized denominator formulation. The routing module processes these aligned tokens through layer normalization, linear projections, and non-linear activation layers, summarizing inter-branch relationships across the current audio sample before applying an explicit uncertainty correction penalty to adjust the normalized gate weights.

The experimental evaluation was conducted on logical-access partitions from the ASVspoof2019 and ASVspoof2021 benchmarks alongside the WaveFake dataset, applying frozen decision thresholds derived from a discrete equal-error-rate rule on development splits. In their evaluation on ASVspoof2019 logical-access data comprising 71,237 evaluation trials, the complete serial system achieved a precision of 99.995 percent and a missing rate of 2.085 percent at an applied threshold of 0.1734. The underlying confusion matrix recorded 62,550 true positives, 3 false positives, 1,332 false negatives, and 7,352 true negatives, yielding a two-sided 95 percent Clopper-Pearson exact confidence interval for precision between 99.986 percent and 99.999 percent, alongside an exact missing-rate interval between 1.976 percent and 2.199 percent.

When the ASVspoof2019-trained model was evaluated on the 148,176 trials of the ASVspoof2021 logical-access evaluation set without target-dataset training or calibration, it attained an observed precision of 99.815 percent and a missing rate of 0.930 percent. This operating point reflected 132,120 true positives, 245 false positives, 1,240 false negatives, and 14,571 true negatives, corresponding to exact 95 percent confidence intervals of 99.790 percent to 99.837 percent for precision and 0.879 percent to 0.983 percent for missing rate. For the WaveFake benchmark, the protocol utilized nonoverlapping within-dataset splits where all four expert detectors and the routing backend were trained and evaluated exclusively on WaveFake data without ASVspoof weights, yielding 26,691 true positives, zero false positives, 2 false negatives, and 2,649 true negatives, which produced an observed precision of 100.000 percent with an exact interval of 99.986 percent to 100.000 percent and a missing rate of 0.007 percent.

Comparative analyses against individual detector branches demonstrated that the complete framework does not universally outperform single backbones across all metrics simultaneously. On ASVspoof2019, each standalone expert branch achieved an observed precision of 99.997 percent, with the spectrogram detector reaching a lower missing rate of 1.091 percent compared to the integrated system’s 2.085 percent. On ASVspoof2021, while the full model improved precision compared to the standalone spectrogram and time-frequency branches, the phoneme detector achieved a higher precision of 99.825 percent, and the time-frequency structure detector recorded a lower missing rate of 0.569 percent. Threshold-independent equal-error-rate assessments recorded 0.177 percent on ASVspoof2019 and 1.229 percent on ASVspoof2021, though the authors emphasize that literature comparisons remain contextual due to differences in training and preprocessing protocols.

Ablation experiments isolating the fusion mechanism revealed that dynamic routing alters operating-point trade-offs rather than guaranteeing uniform error reductions across all settings. When compared against static learned weighting and uniform embedding averaging over identical detector backbones on ASVspoof2019, the dynamic architecture produced lower precision and a higher missing rate than both fixed averaging approaches. On ASVspoof2021, the dynamic router improved observed precision over static weights but exhibited a higher missing rate than simple averaging.

Analysis of aggregate routing distributions highlighted distinct weight patterns across speech categories and corpora. Across both ASVspoof evaluation sets, the routing network allocated its largest mean weight to the time-frequency branch for fake speech samples, averaging 0.660 on ASVspoof2019 and 0.671 on ASVspoof2021, while allocating minimal mean weight to the raw waveform branch. Conversely, for WaveFake synthetic audio, mean routing weights were distributed almost uniformly across all four expert branches, ranging between 0.218 and 0.266. The researchers caution that these pooled sample means illustrate overall network weighting behavior rather than causal explanations or sample-level acoustic correlations.

Computational profiling on an NVIDIA RTX 4090 GPU demonstrated substantial resource requirements for the full multi-branch framework. Evaluating a single 4.0375-second audio input at 16 kHz using full single-precision floating-point execution required 1,292.346 million parameters and registered a mean forward latency of 102.919 milliseconds. This computational cost represents approximately 4.260 times the average forward latency and 4.001 times the parameter count of an individual expert branch. Because the serial implementation evaluates all four backbones sequentially before routing, the system functions as a dense multi-stream network rather than a sparse conditional-computation model, requiring 4.953 GiB of peak allocated GPU memory and an overall checkpoint size of 15.266 GB.

The authors explicitly delineate multiple structural boundaries regarding the interpretation of their findings, noting that the evaluations do not establish performance in noisy acoustic environments, in-the-wild recordings, multilingual speech, or unseen generative speech models. Furthermore, the perfect precision observed on WaveFake reflects a controlled within-dataset experiment rather than zero-shot cross-corpus generalization. The reported exact confidence intervals reflect binomial trial variability conditioned on single trained instances and frozen development thresholds without accounting for acoustic dependencies or repeated-run variance. Consequently, while the findings confirm the feasibility of integrating detector-state variables into sample-dependent routing, they indicate that practical adoption requires balancing operating-point objectives against substantial computational overhead.

Subject of Research: Dynamic routing perception-decision architecture for machine-generated fake speech detection

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

Article References: Luo, X., Wu, X., Xu, M., Hamdulla, A., & Zheng, T. F. (2026). Enhancing the robustness of brain-inspired perception decision machine for fake speech detection. Frontiers in Signal Processing, 6, 1908628. https://doi.org/10.3389/frsip.2026.1908628

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1908628

Keywords: fake speech detection, dynamic routing, perception-decision machine, evidential uncertainty, ASVspoof, WaveFake, acoustic features, audio forensics, deepfake audio, one-class softmax, speech synthesis, voice conversion, neural detectors, decision fusion, equal error rate, benchmark evaluation