Voice-processing systems that perform speaker verification, keyword spotting, speech recognition, etc. need complete access to the speech signal, albeit in parameterized form. These data could potentially be logged for future playback, analysis or even malicious activities and represent a threat to the privacy and security of users. This project aims to develop techniques that enable some key voice processing tasks, namely speaker identification or verification and keyword spotting, while preserving the privacy of the speaker?s voice. The techniques will perform their operations without observing any intelligible form of the speech signal from which one could glean any information about the speaker or what they said; yet at the end of the computation the results, which will only be delivered to an authorized party, will be indistinguishable from those that would be obtained if the system were not secured in this manner.
The proposed work draws upon approaches from cryptography and secure multiparty computation. It is explained how these techniques can be used to devise privacy-preserving algorithms for voice processing, and the development of such algorithms for the three problems mentioned, speaker identification and verification and keyword spotting, has been proposed.
For further information see http://mlsp.cs.cmu.edu/projects/secureaudio
