Extension of voice biometrics to predict health and behavior raises privacy concerns
Voice biometrics are increasingly being proposed and used not just to determine who people are, but for voice analysis to learn other things about them, such as how they are feeling, or what they might do in a given situation, The Verge reports. While such analysis may be valuable for a range of businesses, it also raises serious privacy issues.
Voice is becoming ubiquitous as a method of interaction, with the proliferation of digital personal assistants and smart home devices, and there are several companies making interesting claims about the insights they can pull from speech data. Voicesense, for instance, tells potential customers that it can predict the likelihood of defaulting on a bank loan, buy a more expensive product, or succeed in a new job opportunity with real-time voice analysis.
Voice analysis applies machine learning to speech characteristics such as tone, speed, emphases, and pauses. Typically, judgements are made based on complex sets of features and patterns indiscernible by humans. Researchers have built algorithms that have some success identifying Parkinson’s disease and PTSD, and projects like CompanionMx are seeking to use voice analysis to detect mental health emergencies by having patients record audio diaries in an app.
CompanionMx has been tested with more than 1,500 patients over a seven-year period, and results recently published in the Journal of Medical Internet Research indicate some success predicting depression and PTSD. The product spun out of voice analysis company Cogito, and has received funds from the Defense Advanced Research Projects Agency (DARPA) and the National Institute of Mental Health.
Most voice analysis deployments today are in call centers for improving customer engagement, from provider such as Avaya, which added sentiment analysis to its voice platforms last October. Voicesense CEO Yoav Degani, however, says it could be used for predicting insurance claims, identifying investment style, and assessing employee retention. In a trial with a bank, the company identified customers with low and high risks of defaulting on a loan. Those in the low risk group defaulted 6 percent of the time, and those in the high-risk group did 27 percent of the time.
“We’re not correct 100 percent of the time, but we are correct a very impressive percent of the time,” Degani says. “We can provide predictions about health behavior, working behavior, entertainment, so on and so forth.”
MIT research scientists Satrajit Ghosh warns that he does not take claims for granted without proof they have been validated with enough quantity and diversity of samples. Drexel University Criminology Professor Robert D’Ovidio says that even if they are accurate consumer protection regulation might be needed to prevent voice analysis systems from being used to discriminate, for instance against people applying for a mortgage whose voice reveals a medical condition. The risk of algorithms entrenching or reflecting bias has also been subject of increased discussion over the past year.
Such concerns were recently raised by The Intercept after Amazon patented technology to recognize emotional state, accent, and other characteristics.
The Verge notes that the market for voice technology is forecast by IdTechEx to reach $15.5 billion by 2029.
Comments