Abstract

Unique biological and behavioral characteristics are used for biometric identification because they contain reliable subject dependent information. Although conventional modalities such as facial geometry, vocal tract anatomy, fingerprints, iris patterns, and to some extent; gait also provide discriminative capabilities, their use is limited because of acquisition complexity, invasiveness, and sensitivity to environmental conditions. The growing need of non-contact, privacy-preserving biometric systems, research attention has shifted toward acoustic signals inherently being subject dependent. In this context, the Head Related Transfer Function (HRTF) has proven to be a reliable auditory biometric feature. It is direction and frequency dependent filtering of sound by the head, pinnae, torso, and shoulders. HRTF captures three cues: interaural time difference (ITD), interaural level difference (ILD), and pinna-induced spectral shaping. These spatially dependent cues vary in accordance with the morphological structure of the ear and its surrounding region. Our hypothesis is that HRTF not only encode subject specific information, gender-specific information may also be hidden within the acoustic signatures resulting from anatomical distinctions among the subjects. Cranial geometry, pinna curvatures, and torso volume are responsible for generating subject dependent acoustic signatures in the form of Head Related Impulse Response (HRIR). In the present research, a three-stage framework including spectrographic analysis, parametric cepstral visualization, and deep learning based acoustic signal classification has been used. Spectral and cepstral analyses showed gender dependent trends. Male subjects had stronger low frequency energy distribution and higher spectral variability. Female subjects showed enhanced sensitivity to high frequency content. The HRIR of female subjects had smoother cepstral gradients as compared to that of male subjects. For further verification, a hybrid Convolutional Neural Network with Bidirectional Long Short-Term Memory (CNN-BiLSTM) model was used. The model provided gender classification accuracy above 82%, a mean ROC-AUC of 0.98 and F1-scores above 0.95. As the public HRIR datasets have small number of subjects, Leave-One-Subject-Out (LOSO) cross-validation strategy was also used to ensure complete subject independence between training and testing data. The results shows that HRIR can effectively be used for extracting the gender specific information hidden inside the given transfer functions derived from the corresponding HRTF.

Statistics