To address the challenges in transformer voiceprint detection, namely, signal susceptibility to interference and difficulty in obtaining sufficient training samples, a transformer voiceprint fault diagnosis method that integrates acoustic pattern ridging and meta-learning is proposed. First, based on the ridge feature processing, the optimized transformer voice pattern time spectrum is selected for physical feature screening and morphological feature compression. A selective encoder (SE) is then constructed to enhance the correlation between time-frequency and morphological representations, improving model convergence speed. Finally, a meta-learning network is designed for transformer state evaluation. An OD-Reptile-based first-order gradient update strategy is introduced, and an inner-outer loop optimization mechanism is used to enhance parameter generalization. This enables accurate voiceprint diagnosis under limited sample and noisy conditions. Compared with traditional deep learning methods, such as R-WDCNN, LSTM, and CNN, the proposed method reduces the number of convergence iterations by more than 10 rounds in low-sample, high-noise environments (SNR is -12 dB). Accuracy is improved by 6.35%, 12.1%, and 16.93%, respectively. Experimental results demonstrate significant improvements in accuracy, noise immunity, robustness, and generalization.