Cascaded biometric control system based on speaker and speech recognition

In this thesis, a hybrid software is proposed for a system that can be controlled by speaker recognition and then speech recognition in biometric identification systems. In the first phase of the study, text-independent speaker recognition was performed, after that it is combined with the recognition of the password of the defined user, and a hybrid voice recognition process is revealed. Therefore a voice biometric system that uses speech and speech recognition methods together is presented. This concatenated method reduces the False Positive Rate (FAR) and increases the safety of the biometric recognition system. Mel Frequency Cepstral Coefficient (MFCC) method is used in the first phase of the study, as it is known that it provides the desired attribute contribution for speaker recognition; however, due to the noise-prone behavior of the MFCC, a fast noise suppression approach, the noise gate is used at this point. Since the signal to noise ratio (SNR) is higher in the unpartitioned portions of the speech, these portions of the speech signal are detected using short time energy and interrupted. If the short-term energy of a square is lower than a pre-defined threshold, it is considered noise and this section is removed. The MFCC values obtained are used as biometric templates. The method used when new entry into the target system is based on whether the Euclidian distance metric between the MFCC values of the test data and the MFCCs recorded as templates is greater than a certain threshold value. The voice recognition phase is based on the minimum field frequency obtained from autocorrelation and advanced signal processing. In this thesis study, MATLAB-based software for cascade voice recognition method is presented. Thus, it has been shown that voice recognition is better defined in terms of accuracy, security and penetration difficulty. According to the data obtained in the study, the efficiency of the identification system was approximately 91.2%. Keywords: Speaker recognition, mel frecquency cepstral cofficients, pitch period.