Spoken Pass-Phrase Verification in the i-vector Space

The task of spoken pass-phrase verification is to decide whethera test utterance contains the same phrase as given enrollmentutterances. Beside other applications, pass-phrase verificationcan complement an independent speaker verification subsystemin text-dependent speaker verification. It can also be used forliveness detection by verifying that the user is able to correctlyrespond to a randomly prompted phrase. In this paper, we buildon our previous work on i-vector based text-dependent speakerverification, where we have shown that i-vectors extracted usingphrase specific Hidden Markov Models (HMMs) or using DeepNeural Network (DNN) based bottle-neck (BN) features help toreject utterances with wrong pass-phrases. We apply the samei-vector extraction techniques to the stand-alone task of speakerindependentspoken pass-phrase classification and verification.The experiments on RSR2015 and RedDots databases show thatvery simple scoring techniques (e.g. cosine distance scoring)applied to such i-vectors can provide results superior to thosepreviously published on the same data.