Spies want to disguise voices while researchers work to detect voice spoofs
A U.S. intelligence project incubator has announced it is looking for help creating anonymous real-time speech algorithms that disguise a speaker’s identity live.
Meanwhile, a pair of researchers in China say they have designed biometric code that is better than competing approaches at performing the opposite task – detecting voice spoofs.
The Intelligence Advanced Research Projects Activity agency has posted an invitation to develop code that can mask biometric identifiers in conversational speech. A proposers’ day will be held June 27.
Given the agency involved, anything that comes from the collaboration would likely be used government spycraft than anti-fraud capabilities aimed at consumers. Industry could eventually get a crack at the resulting research, however, given the multiplying electronic goods and services that operate with voice commands.
Specifically, IARPA officials want to see innovations that defeats threats to privacy including speaker identification, human evaluation of static traits and automated classification of dynamic traits.
The best efforts will hide identity while delivering natural-sounding conversation.
A pair of researchers at Zhengzhou University say most voice biometric spoof detection algorithms examine “the physical features of speech” to do their work. They say they can determine a speaker’s gender, age and mouth shape by examining a “large number” of physiological features associated with a human face.
The physiological is combined with the physical and a densely connected convolutional neural network to spot spoofing.
They claim that, in the ASVspoof 2019 logical access scenario, their presentation attack detection algorithm improves the tandem decision cost function and equal error rate by 5 percent and 7 percent when compared to other options available.