Like any other biometric systems, voice authentication systems can be easily spoofed. Making use of synthetic speech samples produced either using advanced text-to-speech or voice conversion technology one can easily create synthetic voices that sounds as natural as if spoken by a real human and therefore it is quite hard for a human ear to distinguish. Therefore, fraudsters can make use of such technology to break into someone else’s voice controlled authentication system. Another effective method that does not require much expertise or knowledge of any technology is replay attack. This form of attack involves just playing back pre-recorded voice samples (stolen voice) to a voice-driven authentication system to gain illegitimate access. And such attack can be easily launched using smart phones from a quite room/environment. Furthermore acquiring samples of digital voice print of target speaker is not hard in today’s age of social media.
Our experts and researchers from R&D put a huge effort in voice anti-spoofing research as we believe it is one major component in voice authentication to enable trust to its user. If there is no trust why would users use it in first place. Therefore, our dedicated R&D team keep upto-date with the latest update in the technology by attending/presenting in top speech conferences and incorporate existing engines with latest technologies. The artefacts and cues present in computer generated speech or replayed speech can be quite hard to detect by human ears as our ears are not designed to detect such speech signals. Our algorithms implement in voice anti-spoofing is capable of detecting cues to a greater accuracy and help end users to fall in wrong hands.
Our voice anti-spoofing solution is built combining many different techniques in a form of ensemble to determine if a voice is a live one. This includes combination of both generative models such as Gaussian Mixture Models, Auto-encoders and discriminative modelling using deep neural networks. Our models work on both hand-crafted features such as Constant-Q-Cepstral Coefficients and linear frequency cepstral coefficients that has proven to show good results in detecting spoofing attacks; and we also employ end-to-end model which works on direct raw waveforms.
Our R&D team also actively participates in the ASVspoof challenge which is a premier anti- spoofing challenge conducted bi-annualy by the Speaker verification community. Visit our R&D page to learn more about our ongoing research work and related publications from the lab.
Please get in touch with us for a free demo.
Voice aunthentication (or person verification) uses automatic speaker recognition technology. The goal here is to automatically recognize/verify the identity of a claimed person using their voice. These systems are deployed in two major settings: text dependent and text-independent settings depending upon the level of user cooperation. In text dependent application, the system has prior knowledge of the spoken text and therefore expects same utterance during deployment. On the contrary, in text independent systems there is no prior knowledge about the lexical contents, and therefore these systems are much more complex than text dependent ones.
Please reach us through the following email if you have any questions.
info@boracsolutions.com