Multimodal Biometric System (Multi-BioTM)
Multimodal biometrics system has received increasing attention in recent years. Theoretical foundation of multimodal biometrics can be found in early work of combining classifiers. Existing multimodal systems include face-fingerprint-hand geometry, face-iris, face-voice and so on. The two modalities we chose to study in this project (fingerprint and voice) support low-cost acquisition applications and more importantly have complementary properties (one has peaked intra-class distribution and the other has peaked inter-class distribution).
The figure below shows the dramatic improvement in personal identification by using fingerprint and voiceprint.
Multimodal fusion dramatically improves the performance due to complementary properties of fingerprint and voice biometrics (WVU database with over 100 subjects).
We also developed a real-time prototype that can perform multimodal fusion. The prototype is standalone, runs in C, is modular and portable.
The software interface is shown below. In order to capture the voiceprint, we used a commercial software called Goldwave. The captured voiceprint is saved in a designated input directory. Our processing software searches the input folder at an interval of 5 seconds. If there is a new file, the software will grab it and process it. For the fingerprint, we used the SerialCom software developed by Fingerprint Cards to capture the fingerprint. Similar to voiceprint, the captured fingerprint is saved at a designated folder. Our processing software scans that folder at an interval of 5 seconds.
The interface of our multimodal biometric system has 4 sub-windows. The top left one shows the voiceprint. The lower left shows the decision results, including scores and matched personís name. The top right one shows the system status and the lower right one shows the processing time of each key module.
The following video clips include three demos: 1) fingerprint only; 2) voiceprint only; 3) fingerprint and voiceprint combined. There are 3 people in the reference database. Each person has enrolled 4 fingerprints: right index, right middle, left index, left middle. The output scores are in that order. The highest score one will be chosen as the decision for fingerprints. Similarly, each person records 4 voiceprints (10 digits for each voiceprint). The lowest score one will be chosen for voiceprint. This difference is caused by the different matching algorithms used in fingerprint and voiceprint recognition. Five detailed demo video files are resided in the Download section.