Voiceprint Identification System For The Applications Of Satisfaction Telephone Interviews Cheating Investigation

In the car market customer satisfaction survey, dealers might change the telephone number of consumer, which cause the obtained survey results not true. The traditional verification methods could not effectively troubleshoot those impostors, In this case, we can use the technology of speaker identification based on biometric technology, which can effectively identify the impostors from the recording data of telephone interviews. Due to the client’s telephone equipment and transmission session may not identical, the recording and session variability problem during speaker identification will be more complicated. Another problem is that some manufacturer’s original recordings are mixed with two-side voices. To conquer these two problems and improve the system’s identification performance, this article will focus on the study of intersession and speaker diarization.We adopt a series of effective algorithms, including feature domain, model domain and score domain, to reduce the harmful effect of intersession. In the feature domain, we apply the method of feature warping. In the model domain, we combine SVM with NAP, and LFA with UBM-GMM. In the score domain, we use T-norm to further improve the performance.For the segmentation and clustering of speaker diarization, we adopt the adaptive GMM based on UBM, to solve the problem of lack priori knowledge. And use the BIC method to optimize segmentation results in the initial splitting step. Speaker segmentation and clustering is one of the main factors affect the performance of cheating investigation system. The segmented results will directly affect the system final recognition rate.Based on the above two issues, we have successfully built the cheating investigation system for the telephone interview of automotive customer satisfaction.