Voiceprint detection technology can assist inspectors in assessing the state of transformers. A method for detecting and assessing transformer states based on voiceprint compression and cost-sensitive techniques is proposed. The method first extracts voiceprint features from transformer audio,then filters and compresses these features in the frequency domain. Subsequently,a convolutional neural network is employed to evaluate the transformer's state,incorporating a cost-sensitive loss function to enhance attention towards difficult samples. Using a 35 kV transformer as the experimental subject,transformer audio data is collected through on-site recordings,simulated experiments and sample augmentation. Test results demonstrate that the proposed method reduces the voiceprint dimensionality from 1 025 to 80,decreasing computational complexity and video memory usage to 8.1% and 7.7% of the original 1 025 dimensions,respectively. Simultaneously,the proposed method achieves a voiceprint recognition accuracy of 83.5% and improves the recall rate of the most challenging short-circuit current anomaly from 48.2% to 63.6%.