JFA for text-dependent speaker verification

点击翻译按钮获取中文摘要

VoiceTrust needs to be able to authenticate speakers using utterances of short duration (1-2 seconds). Their applications require a technology which is robust to channel effects (so that, for example, speakers can use different handsets at enrollment and verification time). Furthermore this robustness needs to be achieved using modest amounts of data for classifier training. To satisfy the duration constraint, speaker verification needs to be ``text-dependent'' (that is, based on passphrases). This precludes the use of state of the art methods of text-independent methods of speaker verification and, in particular, the i-vector/PLDA approach. We propose to satisfy VoiceTrust's requirements by developing a version of Joint Factor Analysis (JFA, the predecessor of i-vector/PLDA) which is tailored to the text-dependent speaker recognition problem and which can be trained with modest amounts of data. We will develop two solutions to the problem, one designed to handle the case where background data for modeling passphrases is available and the other to handle the general case. (The technical challenge is greater in the general case but success here would relieve VoiceTrust of the burden of collecting data.) To achieve robustness to additive background noise and reverberation we will develop several acoustic feature extractors and train a JFA back end in each case (so that there will be one classifier for each front end). Scores produced by these classifiers will be fused by logistic regression to make speaker verification decisions. Finally VoiceTrust is particularly concerned that speaker recognition technology is vulnerable to spoofing attacks by speech synthesis and voice conversion technology. We will create synthetic datasets and use them to train JFA-like statistical models of synthesis and voice conversion artifacts in order to detect spoofing attacks. This technology will be integrated into currently existing VoiceTrust products, VT-Athena and VT-Aura.