An optimization method of Voiceprint Recognition based on user portrait

Voiceprint is an important component of creating a user portrait. Voiceprint Recognition can determine user's identification. However, speech signals in the customer service system are processed by encoded with compression for effective transmission and storage. The low-bit rate codec results that the performance of Voiceprint Recognition system dramatically reduces. What is more, the speech number of each customer is not adequate. In order to solve the problem, this paper proposes a model compensation method. The method uses a test utterance with expectation maximization (EM) algorithm to estimate the distortion model and the UBM is adjusted to match the codec type of the test utterance. Voiceprint Recognition experiments are conducted. The results show that the proposed method is able to dramatically improve the performance of the system.