Adding voice to facial recognition can help in finding important details in child abuse videos

Australian researchers say it is possible to add another biometric to improve results in the depressingly monumental task of enforcing online child sex abuse laws — voice.

A recent paper from Australia’s Institute of Criminology describes a prototype method of identifying both face and voice biometrics. Facial recognition systems can be powerful tools, but as any criminal prosecutor will say, more evidence means more convictions.

Getting face and voice prints from the same content can better identify victims and perpetrators, especially where a face is obscured.

Researchers were working with clips that were between 10 seconds and 12 minutes, and averaging 118 seconds.

Voice recognition in the prototype was achieved first by using a voice activity detector to find segments that contain only voices, making biometric sample matching easier. Work can still be done with a noisy segment using specialized algorithms, however.

Just such an algorithm was used by the Australian researchers, code published in 2020 by Chowdhury and Ross, according to the paper’s authors. A convolutional neural network fuses the output.

The algorithm compares segments to those of all other collected clips.

Researchers created a new interface, through which to view the face and voice matching, enables operators to adjust match scores to meet needs presented by content.

According to the paper, “any observed deficiencies with face recognition … and speaker recognition … are overcome by combining these biometric attributes.”

The U.S. Department of Homeland Services Science and Technology Directorate launched a program to identify victims of child abuse videos on the dark web with face biometrics back in 2018.

Article Topics

biometric identification  |  biometric matching  |  biometrics  |  biometrics research  |  criminal ID  |  facial recognition  |  forensics  |  voice biometrics  |  voice recognition