WhisperNet: Deep Siamese Network For Emotion and Speech Tempo Invariant Visual-Only Lip-Based Biometric

• Published in: 2021 7th International Conference on Signal Processing and Intelligent Systems (ICSPIS) pp. 1 - 5
• Main Authors: Zakeri, Abdollah, Hassanpour, Hamid
• Format: Conference Proceeding
• Language: English
• Published: IEEE 29.12.2021
• Subjects: Authentication; Biometrics; Biometrics (access control); Deep learning; Deep Siamese Network; Face recognition; Lip-Based Biometrics; Privacy; Signal processing; Video Processing; Visualization
• DOI: 10.1109/ICSPIS54653.2021.9729394

In the recent decade, the field of biometrics was revolutionized thanks to the rise of deep learning. Many improvements were done on old biometric methods which reduced the security concerns. Before biometric people verification methods like facial recognition, an imposter could access people's vital information simply by finding out their password via installing a key-logger on their system. Thanks to deep learning, safer biometric approaches to person verification and person re-identification like visual authentication and audio-visual authentication were made possible and applicable on many devices like smartphones and laptops. Unfortunately, facial recognition is considered to be a threat to personal privacy by some people. Additionally, biometric methods that use the audio modality are not always applicable due to reasons like audio noise present in the environment. Lip-based biometric authentication (LBBA) is the process of authenticating a person using a video of their lips' movement while talking. In order to solve the mentioned concerns about other biometric authentication methods, we can use a visual-only LBBA method. Since people might have different emotional states that could potentially affect their utterance and speech tempo, the audio-only LBBA method must be able to produce an emotional and speech tempo invariant embedding of the input utterance video. In this article, we proposed a network inspired by the Siamese architecture that learned to produce emotion and speech tempo invariant representations of the input utterance videos. In order to train and test our proposed network, we used the CREMA-D dataset and achieved 95.41 % accuracy on the validation set.