Abstract: Diagnosing anomalous power transformer voiceprints requires extracting discriminative features under complex, high-noise operating conditions. Existing deep-learning methods for anomaly diagnosis using power transformer voiceprints suffer from insufficient robustness under long-tailed data distributions and underutilization of phase information. To address these issues, this paper proposes a recognition method that integrates dual-spectrogram features with a dual-stream Single-stage Fusion Mamba. First, a feature front-end is constructed in which a log-Mel amplitude spectrogram and a modified group-delay phase spectrogram are mutually complementary, jointly characterizing anomalous voiceprints from both amplitude and phase perspectives. Second, a Cross-Gated Frequency Refinement (CGFR) module is designed, which performs cross-gated modulation between two complementary branches—spatial and channel—to adaptively select cross-branch information and applies frequency-domain filtering to enhance global consistency, thereby purifying time-frequency features and highlighting fault-discriminative regions. A Single-stage Fusion Mamba backbone then performs single-stage cross-spectral fusion to complete anomaly diagnosis. During training, a class-balanced cross-entropy loss based on the effective number of samples is employed to mitigate extreme class imbalance, and stratified five-fold cross-validation is used for confidence assessment. With 5.76 M parameters, the proposed method achieves a macro-averaged F1 score of 96.77% and a tail-class recall of 94.00% under −5 dB strong noise. The results demonstrate that the method effectively improves the robustness of power transformer anomaly recognition—especially for rare faults—under high-noise operating conditions.