Abstract: Voiceprint models based on the closed set are trained with several kinds of fault samples and normal samples. Meanwhile, it is assumed that the subsequent input samples will belong to these several kinds or sets. In practice, strong environmental noises which are unfamiliar to the models will cause a lot of false alarms. In order to reduce the false alarm rate in noisy environments and keep the recall of the fault samples, this paper proposes a two-stage voiceprint recognition algorithm. The stored energy motors of the circuit breakers were taken as the research object. The sound signals from a normal working condition and three kinds of abnormal working conditions were captured and preprocessed with framing and windowing. Then, the Mel frequency cestrum coefficient (MFCC) feature vectors were extracted. Finally, the One-class support vector machine (SVM), a novelty detection algorithm, was used as the first stage algorithm for the separation of the samples polluted by noise, and the C-SVM as the second stage algorithm for the recognition. The experiment results show that compared to the existing single-stage voiceprint recognition algorithms, this method can effectively separate the polluted samples and reduce the false alarm rate by over 63.47% when the signal-to-noise ratio (SNR) is lower than -20dB and the decrease of fault samples' recall keeps lower than 0.33% when the SNR is higher than 20dB.