TY - GEN
T1 - Evaluation of Bone-Conducted Cross-Talk Sound in the Head for Biometric Identification
AU - Irwansyah,
AU - Otsuka, Sho
AU - Nakagawa, Seiji
N1 - Publisher Copyright:
© 2021 ACM.
PY - 2021/10/5
Y1 - 2021/10/5
N2 - Just like a fingerprint, every human head has its own uniqueness. To take advantage of its unique geometrical and physical characteristics, we can insert an inward-facing microphone into the ear canal to record sound coming from a bone conduction transducer on the mastoid on the contralateral side of the head. This sound is known as "cross-talk"sound, which is unique for each individual. In this study, we present an evaluation of bone-conducted "cross-talk"sound in the head for biometric user identification. Our approach relies on "cross-talk"sounds to estimate impulse responses (IRs) and uses them to identify the corresponding users. Mel frequency cepstral coefficients (MFCCs) are extracted from a 16-ms IR as acoustic features, and 1-nearest neighbor (1NN) classifier is used for making a decision. Finally, we evaluated the proposed approach with ten participants. Our results showed that "cross-talk"sounds could be used to identify users with an average accuracy of up to 99.8%, and the equal error rate (EER) obtained was 2.6% in user authentication. In addition, the IRs dataset and a video demonstrating how the system worked were made available on GitHub and YouTube.
AB - Just like a fingerprint, every human head has its own uniqueness. To take advantage of its unique geometrical and physical characteristics, we can insert an inward-facing microphone into the ear canal to record sound coming from a bone conduction transducer on the mastoid on the contralateral side of the head. This sound is known as "cross-talk"sound, which is unique for each individual. In this study, we present an evaluation of bone-conducted "cross-talk"sound in the head for biometric user identification. Our approach relies on "cross-talk"sounds to estimate impulse responses (IRs) and uses them to identify the corresponding users. Mel frequency cepstral coefficients (MFCCs) are extracted from a 16-ms IR as acoustic features, and 1-nearest neighbor (1NN) classifier is used for making a decision. Finally, we evaluated the proposed approach with ten participants. Our results showed that "cross-talk"sounds could be used to identify users with an average accuracy of up to 99.8%, and the equal error rate (EER) obtained was 2.6% in user authentication. In addition, the IRs dataset and a video demonstrating how the system worked were made available on GitHub and YouTube.
KW - biometrics
KW - bone conduction
KW - cross-talk
KW - user identification
UR - https://www.scopus.com/pages/publications/85125425866
U2 - 10.1145/3489088.3489119
DO - 10.1145/3489088.3489119
M3 - Conference contribution
AN - SCOPUS:85125425866
T3 - ACM International Conference Proceeding Series
SP - 76
EP - 80
BT - Proceedings of the 2021 International Conference on Computer, Control, Informatics and Its Applications - Learning Experience
PB - Association for Computing Machinery
T2 - 2021 International Conference on Computer, Control, Informatics and Its Applications - Learning Experience: Raising and Leveraging the Digital Technologies During the COVID-19 Pandemic, IC3INA
Y2 - 5 October 2021 through 7 October 2021
ER -