Abstract: This paper proposes a modeling and identification algorithm of recording devices. The features of “the fingerprints of recording devices” are extracted from the silence periods of a speech for the purpose of recording devices identification, because the same device information can be obtained from both silence and speech periods in an audio sequence but the device information is not effected by speaker information, texture information and emotion information in the silence periods. On the one hand, the recording device universal background model is used to establish the model of inverse devices. On the other hand, the model of specific recording device is achieved by adapting the recording device universal background model through MAP adaptation algorithm. Finally, the normalized log-likelihood is used to classify the specific categories of the recording devices of the input speech samples. The results of experiments indicate that the average accuracy of recording device identification on 9 recording devices is 87.42% and the effects of different factors on the impact of the proposed algorithm are investigated, which proves the effectiveness and reliability of the proposed algorithm.