Analysing encrypted speech

Abstract

Speech is digitized, encrypted and sent between two parties in many situations. Digitized speech signals are generally considered as ordinary binary data streams as far as encryption is concerned. In this thesis, we present that the properties of speech signals are different from text, image, video and other non speech signals and need special attention while being encrypted. These properties of speech signals have to be kept in mind, not only, while designing encryption algorithms for digitized speech signals but also during implementation of these algorithms in software and hardware. We present that how the statistical properties of speech can be utilized to extract important information, from cryptanalytic point of view, from the encrypted speech signals. We also study some of the published cryptanalysis techniques particulary used for text based data and then study the effectiveness of these techniques if the underlying plaintext data is digitized and compressed speech. We discuss a particular situation in stream ciphers called the two time pad situation, assuming the two underlying signals being speech. The two time pad situation is presented at length in the literature with respect to text based data and not discussed at all with respect to secure speech communications. We present techniques for exploitation of two time pad situation in case of stream ciphered digitized speech. Separate techniques for exploitation are presented for waveform (uncompressed) and parameter encoded (compressed) speech. We also analyze the latest techniques of selective encryption applied on compressed speech signals and present that some of these techniques can be effectively employed for content protection but may not conceal the speaker’s identity. In the last part of the thesis we analyze the latest encryption techniques used in voice over IP applications. We discuss that if speech is compressed using variable bit rate encoding resulting into variable length packets can be used for speakers recognition even after encryption. Experimental evaluation results for all the above techniques are presented.