| Summary: | 碩士 === 中央警察大學 === 刑事警察研究所 === 106 === Speech and signature are two major ways which frequently used by humans in daily trading and interacting, they are contactless and quite easily to get, which extremely suitable for on-line verification. This essay integrates speaker verification with signature verification and establishes a verification system with multi-feature. We use Mel frequency cepstral coefficients (MFCCs) as the voice characteristic for speaker verification, and the voice experiment is text independent. The characteristic of signature verification is the projection of horizontal projection on Y-axis and vertical projection on X-axis and the experiment of signature is on-line system. The algorithm uses modified Gaussian mixture model, in comparison with gaussian mixture model, it dramatically reduces calculation amount while maintaining kind effects of verification. In this paper, we do five types of experiments in total, there are speaker verification, signature verification, unequal-weights combine speaker verification and signature verification, equal-weights combine speaker verification and signature verification, and combine with speaker verification and signature verification. The result shows the average error rate is 2.552578%, 5.156821%, 0.494697%, 0.794678% and 1.940667%. The above experimental results show that using multi-feature can reduce the error rate of verification and the verification of the above three kind of multi-feature, unequal-weights error rate is the lowest.
|
|---|
