Customer Identification Through Automated Voice Recognition

Abstract

The prevailing practice in the state banks with regard to drawing pension funds by a senior citizen involves the pensioner is required to be physically present at the bank. In order to overcome the above stated hindrance, biometrics technology is introduced by the researcher. The objective of the proposed system is to recognize the customer through his/her voice. Customer recognition by voice is called speaker recognition. Speaker recognition can be categorized into speaker identification and verification. Since we aredealing with known speakers, this research is mainly focused on speaker verification. Speaker identification and verification can further be categorized into text-dependent and text-independent. If the speaker recognition system is text dependent then the system is aware of the text spoken by the speaker, but in the other hand the system should be able to identify the speaker through any text spoken by the speaker. During this research, we have focused on developing a text-independent system, which can increase the usability and security. Four main steps are involved in speaker recognition. The first one is to identify the useful features of voice which makes the speaker s speech unique. Second step is to extract those features by using a feature extraction tool. Next step is to model the speaker by using these features. Finally the verification and decision making steps take place. Selecting an efficient feature extraction tool was a challenge. Mel Frequency Cepstral Coefficients (MFCC) and Linear Predictive Coefficients (LPC) are popular feature extraction techniques. We have selected MELCEPST as the feature extraction tool. Mean value of the elements in voice model is used as feature matching technique. We have built the speech database using different languages like English, Sinhala and Tamil with a few no of speakers and the system was trained and tested with those voices. The MIT commercial corpus was used for the critical evaluation. According to the results obtained from collected data, the threshold value is set as 0.33 and MIT corpus threshold value is set as 0.45. We have implemented the uniqueness of a person s voice successfully and developed customer verification application.