P 3 S 2 : practical secure protocol for speech data publishing.

Speech data publishing discloses users' data privacy, and thus entails more privacy risks for users. Existing work sanitized the content, voice, and, voiceprint of speech data without considering the consistence among these three aspects, and therefore cannot protect users' data privacy. To this end, we propose a practical secure protocol for speech data publishing P3S2, the first attempt towards taking the corrections among the three factors into consideration when it sanitizes users' speech data. To concrete, it designs a three-dimension sanitization that utilizes feature learning to capture the set of characteristics in each dimension, and then sanitizes speech data in each dimension using the learned features. As a result, the correlations among the three dimensions of the sanitized speech data are guaranteed. Furthermore, it utilizes two real world datasets, TED talks and LibriSpeech to evaluate the performance of P3S2 in terms of the data privacy preservation.