원문정보
초록
영어
With the popularity of high fidelity and portable recording device, it becomes very easy for the attacker to spoof speaker verification system by voice playback. In this paper, it found that there exists obvious difference in the scale-factors, a parameter of MP3 codec, when the original and playback voices are compressed by MP3. So, a detection algorithm for the playback voice is proposed. The experimental results show that the detection accuracy of the proposed algorithm can reach to 99.51% for 4 different types of eavesdropping devices. Meanwhile, the algorithm is integrated into the speaker verification system based on GMM-UBM. The equal error rate (EER) of the system has dropped 32% and the system's ability to resist playback attack is improved.
목차
1. Introduction
2. Mathematical Model of Voice Playback Attack
3. Feature Extraction and Verification
3.1. Feature Extraction
3.2. Verification with PSDM
4. Database Set
4.1 Corpus Set
4.2 Playback Speech Database set
5. Experimental Results
5.1 Influences of the Different Classifiers
5.2 Influences of the Eavesdropping Devices' Coding Format
5.3 Verification System with PSDM
6. Conclusions
References