Transformer Voiceprint Recognition Based on the LSTM-CNN-CBAM Model

Existing voiceprint recognition models have limitations in feature extraction capabilities, making it difficult to fully integrate deep semantic information with multi-scale temporal features. A hybrid model based on LSTM-CNN-CBAM is proposed. The voiceprint signal is processed by extracting Mel-frequency cepstrum coefficient (MFCC) features and applying wavelet denoising technology to construct a dataset, which effectively suppresses noise interference and enhances feature representation capability. By introducing the attention module (CBAM) into the LSTM-CNN network, a deep learning model for spatio-temporal feature fusion is constructed, achieving efficient recognition and classification of 8 typical defects such as partial discharge and DC bias magnetization in the voiceprint sequence, as well as the normal state. Experiments show that the model achieves an average accuracy of 93.75% on 9 datasets. Compared with the ResNet18 and LSTM models, the accuracy is improved by 4.09% and 6.31%, respectively. The introduction of the attention mechanism provides a solution with higher accuracy and stronger robustness for transformer voiceprint recognition, which has significant engineering application value.