Speaker recognition implementation for authentication using filtered MFCC — VQ and a thresholding method

This paper explains about authentication mechanism using one of the unique biometric component, the human voice. First of all, the characteristic of the voice is extracted using MFCC then represented by cepstrum coefficients. Later, those features forms a model by the VQ method. These methods are modified with a proposed thresholding method to reject the unknown voice and a Butterworth Filter to handle the noise. For the experiment, we used both synthetic and real human voice, or biometric data. Both synthetic and biometric data consist of 10 speaker. Half of the speaker is separated as the unregistered or the untrained voices. Overall, the result shows that the methods is adequate enough to perform a security mechanism. MFCC and VQ combination can truly 100% distinguish the speakers in a closed sample which includes only the registered speaker. Compared to the noise-added data, the noise-filtered data can increase the true acceptance accuracy with a specific filter parameters. The proposed thresholding method is effective enough to reject the unknown voice with approximately 90% true rejection but produces only around 70% true acceptance. Hence, the value of the threshold tolerance, which is to increase the authentication accuracy for the registered speaker, needs to be treated more for the next experiment to find the balance between the acceptance, and the rejection accuracy.