Parameter optimization in a text-dependent cryptographic-speech-key generation task
Chapter in Scopus
- Additional Document Info
- View All
In this paper an improvement in the generation of the crypto-graphic- speech-key by optimising the number of parameters is presented. It involves the selection of the number of dimensions with the best performance for each of the phonemes. First, the Mel frequency cepstral coefficients, (first and second derivatives) of the speech signal are calculated. Then, an Automatic Speech Recogniser, which models are previously trained, is used to detect the phoneme limits in the speech utterance. Afterwards, the feature vectors are built using both the phoneme-speech models and the information obtained from the phoneme segmentation. Finally, the Support Vector Machines classifier, relying on an RBF kernel, computes the cryptographic key. By optimising the number of parameters our results show an improvement of 19.88%, 17.08%, 14.91% for 10, 20 and 30 speakers respectively, employing the YOHO database. © Springer-Verlag Berlin Heidelberg 2005.