A Note on Biometrics and Voice Print: Voice-Signal Feature Selection and Extraction – A Burg-Töeplitz Approach Khalid Saeed # # Faculty of Computer Science, Bialystok Technical University, Wiejska 45A, 15-351 Bialystok, Poland. aida@ii.pb.bialystok.pl http://aragorn.pb.bialystok.pl/~zspinfo/ Abstract: This work presents new applications of Töeplitz matrix eigenvalues approach in image description, feature extraction and recognition. It discusses the possibility of treating the speech signal graphically in order to extract the essential image features as a basic step in successful data mining applications in the biometric techniques. The considered object here is the human-voice signal. The suggested frequency spectral estimation and Töeplitz-based approach, built on linear predictive coding principle, has proved the possibility of selecting signal features from the power spectral plot and entering Töeplitz matrix in a manner similar to its application on images of written texts, signature, palm-print, face geometry or finger- prints, the topics that have shown a success rate of about 98% in many cases. The extracted feature-carrying image comprises the elements of Töeplitz matrices to consecutively compute their minimal eigenvalues and introduce a set of feature vectors within a class of voices. The required computations were performed in MATLAB proving speech-signal image recognition in a simple and easy-to-use way. This stems from the fact that the presented problem solution and its Matlab implementation do not require to implement any special hardware and can be used in tandem with other biometric technologies in hybrid systems for multi-factor verification. 1 Introduction Voice identification in terms of both speech and speaker authentication has its unique significance and role in almost all known biometric techniques. The reason is simply that people seek for an easy-to-use, reliable and safe system to show and prove their authenticity whenever needed. Each biometric type of technology demands a user to play their active part in the process of being identified and authenticated. They involve user’s writing on paper, stamping their fingerprints, showing their open eye to a camera, pressing their hand to show its geometry, and the most common way of identifying the user by their recalling a code or a password to enter the identifying machine with. Voice biometric methods of human identification, however, need nothing but the human utterance to obtain their voiceprint. The term “voiceprint” was coined as everybody has their own unique and characteristic voice parallel to the so called fingerprint left by fingers of an individual. This term is used in most voice biometric solutions as a template of our unique voice features manifested while entering when entering the identifying system. In this paper, the author introduces some examples and experiments to show how this voiceprint looks like from the graphical point of view and how it is identified for recognition. The basic idea is derived from applying Töeplitz matrix minimal eigenvalues algorithm [1] to Burg's model [2]. This implies a graphical approach for feature extraction, selection and hence signal-image description confronting the conventional and traditional methods. Töeplitz matrix approach is employed to verify a variety of biometrics, including the recognition of hand and machine written texts [3], off [4] and on-line [5] signature, face [6], and voice [7]. In all, it has proved a promising success rate. The same algorithm has also shown its possible application in hybrid systems [3] where multiple forms of classifying and identifying tools are fused in one system. The image of a voice signal in any of its classical forms is rather complicated and usually does not convey exactly similar images of the same signals, even when spoken by the same person. However, Burg's model [2] inspired the author and his team to undertake a promising area of study, and they have managed to discover some new facts. These facts concern the possibility of looking at the voice-signal image in a manner similar to any other object image. This enabled extending Töeplitz matrix applications to cover speech signal description, as well. The experiments and their results in this work have either been published [8] or are under publishing [7]. All of the experiments were performed in Matlab.