Robust speech hashing

Abstract

In this thesis, a robust speech hashing algorithm is proposed and performance of this speech hashing algorithm is compared with several robust audio hashing algorithms. We use phone based frequency-time domain analysis for developing a ngerprint( hash value) for any speech data. Robust Speech Hashing can qualitatively be stated as a \dimensionality reduction" mechanism (which would be called the \robust speech hashing" function) via which the desired content of interest can be tracked and found reliably. Phonemes as speech characteristics and randomized frequency as hashing backbone is used so as to conclude on a secure speech tracking. The proposed algorithm is formed by 3 basic stages: Offline, Online and Comparison stages are applied in order. First, we extract most e ective letter patterns in the cepstral domain. After transforming the speech signal into the spectral domain, the cepstrum coe cients are projected on the subspace spanned by the pattern that represents the letter (vowel) at hand. Moreover a pseudo-random linear transformation is applied in order to add a secure aspect. Lastly, the robust hash values of audio les are compared in the L2 sense. The comparison takes place between between di erent audios as well as same but attacked ones. Several comparison tests are made for robust speech hash value based identi cation. ROC curves for di erent kind of attacks are investigated and we determined that, for speech signals, the proposed algorithm is superior to other considered robust audio hashing functions.