Evaluation of Privacy Protection Techniques for Speech Signals

A ubiquitous networked society, in which all electronic equipment including “sensors” are connected to a network and are able to communicate with one another to share information, will shortly become a reality. Although sensor information is most important in such a network, it does include a large amount of privacy information and therefore it is preferable not to send raw information across the network. In this paper, we focus on privacy protection for speech, where privacy information in speech is defined as the “speaker’s characteristics” and “linguistic privacy information.” We set out to protect privacy information by using “voice conversion” and “deletion of privacy linguistic information from the results of speech recognition.” However, since speech recognition technology is not robust enough in real environments, “speech elimination” technique is also considered. In this paper, we focus mainly on the evaluation of speech elimination and voice conversion.