| Summary: | 碩士 === 中央警察大學 === 刑事警察研究所 === 106 === Gaussian distribution is used in the natural and social sciences to represent real-valued random variables which distributions are not known. Gaussian Mixture Model is a weighted sum of single Gaussian probability density function. It is usually used in various areas of pattern recognition. In this paper, we design programs to extract MFCCs as characteristic parameters of “speaker verification”, and extract numbers of call, in/out ratio, numbers of base station from call data as characteristic parameters of “fraudulent call verification”. Based on Gaussian Mixture Model, this paper presents modified Gaussian mixture model and Gaussian integral model. Since the process of these models without iteration, it can significantly reduce the amount of calculation. And these models still maintain good recognition results. In “speaker verification”, the average error rate on Gaussian mixture model is 0.2083%, on modified Gaussian mixture model is 1.7542%, on normalized modified Gaussian mixture model is 1.7876%. The gap between them is not significant. However, the amount of calculation of modified Gaussian mixture model is much less than Gaussian mixture model. In “fraudulent call verification”, the average error rate on Gaussian mixture model is 28.2200%, on modified Gaussian mixture model is 24.5266%, on Gaussian integral model is 8.2292%. The result of Gaussian integral model is the best among them. In summary, each of modified Gaussian mixture model and Gaussian integral model has a good recognition result in a specific areas. And the amount of cal-culation of them is much less than Gaussian mixture model. These two models can be good algorithm for the speaker verification and the fraud-ulent call verification respectively.
|
|---|
