Discrimination of Pre-recorded Speech and On-site Speech

Autor:	Shao-Qi Zhang, 張紹祺
Rok vydání:	2018
Druh dokumentu:	學位論文 ; thesis
Popis:	106 This study proposes methods for discriminating pre-recorded speech and on-site speech, where the pre-recorded speech means the sound played back from a speaker. For high quality speakers, playing pre-recorded speech would sound like on-site speech, and hence an impostor may attack a voice fingerprint-based security system by using pre-recorded speech. This study would help to avoid such kind of attack. The proposed methods include Gaussian Mixture Model (GMM) classifier and i-vector approach, both of them use the Mel-Frequency Cepstral Coefficients as input features. Our experiments conducted using sixteen subjects voice data and four speakers for playing back purpose show that and i-vector approach performs better than GMM classifier, which attains accuracy of 79.11%.
Databáze:	Networked Digital Library of Theses & Dissertations
Externí odkaz:	http://ndltd.ncl.edu.tw/handle/9nfhqs Zobrazit plný text záznamu