Research On Key Technologies Of Fast Speaker Recognition For Spam Call Filtering

In recent years, with the diversified combination of telecommunication network, Internet and television operators, and along with the steady progress of voice transmission and exchange technology, spam calls have become an increasing threat to the security of people's property as well as social stability and harmony. Therefore, it is now more of common concern to conduct intensive studies on how to safeguard and filter those threatening spam calls. However, the detection of spam calls remains a challenging problem since traditional filtering model is still unable to filter two main types of spam calls, that is, those man-made calls and also those calls whose identity, including phone number and ID, is constantly changing. To address this situation, speaker recognition technology can be used. Speaker recognition technology is closely linked to the aspect of caller and independent of the inherent external factors. In this way, it can effectively prevent spam-call-senders from constantly changing their identity and acting recklessly in communication network. Although speaker recognition technology has been well studied, it is not widely applied in spam call filtering. Given the fact that spam call filtering requires real-time recognition and a fair degree of accuracy, there exist two central problems in the current speaker recognition technology, one is how to choose an appropriate recognition model for spam-voice-senders, and the other is how to achieve rapid recognition on the required number of speakers for spam call filtering.The paper compares different recognition models and several fast speaker recognition methods, and then introduces the combined approach of KLSH and supervector into the area of rapid recognition of spam-call-senders. In order to improve the speed and accuracy of KLSH, a rapid recognition method is proposed in this paper based on the feature space analysis of the speakers in the database, which means using spectral clustering and UBM to select the samples of KLSH. During the process of spectral clustering, the paper focuses on analyzing the similarity measurement method, not only introducing the KL distance but also proposing a similarity measurement method based on the distance of model information entropy. Moreover, block matrix is used to simplify the distance calculation of large-scale speaker models, while the method of automatically acquiring cluster number based on eigen-gap is greatly improved. Next, the GMM supervector is selected as the recognition model of the speakers after comparing the advantages and disadvantages of different supervectors. In addition, the kernel function and approximate search involved in KLSH are carefully analyzed and selected to build a complete KLSH fast recognition system.Experimental results show that, when the test speech is 4s, the average recognition time is 0.105 s and the recognition rate can reach 86.4%, while when the test length is 10 s, the recognition rate is up to 96.7% and the recognition time is only 0.234 s. Compared to the GMM-UBM, it achieves a speed of 1082.4 with the recognition rate only 1.9 percent lower. All the results verify that the algorithm proposed in this paper can be competent for large-scale spam call filtering. At the end of this paper, it is proved that the samples selected by feature space analysis of the speakers have the advantage of certain stability, which means the recognition rate of the original test speakers will not change substantially with a small increase of the speakers in the database.