Detection of synthetic speech for the problem of imposture

Abstract

In this paper, we present new results from our research into the vulnerability of a speaker verification (SV) system to synthetic speech. We use a HMM-based speech synthesizer, which creates synthetic speech for a targeted speaker through adaptation of a background model and both GMM-UBM and support vector machine (SVM) SV systems. Using 283 speakers from the Wall-Street Journal (WSJ) corpus, our SV systems have a 0.35% EER. When the systems are tested with synthetic speech generated from speaker models derived from the WSJ journal corpus, over 91% of the matched claims are accepted. We propose the use of relative phase shift (RPS) in order to detect synthetic speech and develop a GMM-based synthetic speech classifier (SSC). Using the SSC, we are able to correctly classify human speech in 95% of tests and synthetic speech in 88% of tests thus significantly reducing the vulnerability.

Spoofing countermeasures to protect automatic speaker verification from voice conversion

“…In consequence, the approaches to detect converted voice and synthesized speech reported in [15][16][17][18][19][20] will not detect speech converted according to the approach described above.…”

“…As such, converted speech retains real-speech phase information. Attacks of this nature will thus have the potential to overcome the countermeasures proposed in [15][16][17][18][19][20]. This paper reports a new countermeasure which exploits the reduction in pair-wise distances between consecutive feature vectors when they are both shifted towards the same local maxima of the likelihood function of a target speaker model as a consequence of voice conversion.…”

“…Prior work to develop spoofing countermeasures has focused mostly on those aspects of the speech signal not used for recognition. Several works have demonstrated increased robustness to spoofing stemming from speech synthesis through the use of prosodic features and phase [15][16][17]. Based on the This work was partially supported by the TABULA RASA project funded under the 7th Framework Programme of the European Union (EU) (grant agreement number 257289), the ALIAS project (AAL-2009-2-049 -co-funded by the EC, the French ANR and the German BMBF) and by a Futur et Ruptures award from Institut TELECOM. assumption that phase information is generally lost during voice conversion, the utility of phase-related features to detect voice conversion was investigated in [18,19].…”

“…face recognition [12][13][14], the speaker recognition community has started to address the problem through efforts to develop specific spoofing countermeasures [15][16][17][18][19][20]. Prior work to develop spoofing countermeasures has focused mostly on those aspects of the speech signal not used for recognition.…”

Increasing anti-spoofing protection in speaker verification using linear prediction

“…For example, one group of such methods exploits the fact that many speech synthesis and voice conversion algorithms disturb the natural phase of the speech signal. In [9] the authors challenged GMM-UBM and SVM-GMM speaker verification systems with genuine and synthesized speech originating from the WSJ corpus. They showed that by using relative phase shift (RPS) features it was possible to decrease the EER from over 81 % to less than 3 %.…”