University researchers say voice biometrics can tell identical twins apart
Muscle use, emotion and psychological intent lead to differences between identical twins detectable by anti-spoofing voice biometrics tools according to a yet-unpublished study by Ontario Tech University shared with Biometric Update.
The university’s Faculty of Business and Information Technology collected voice samples from two pairs of identical twins in a controlled environment then used voice biometrics tools to analyze the twins’ voice recordings in order to determine the accuracy in detecting imposter attacks. The tools assessed were voice recognition technologies from VoiceVantage (a THC Technologies Corporation subsidiary), Verint, and Microsoft Azure. THC sponsored the research.
From an anatomical perspective, the dimensions of the vocal tract and larynx are responsible for an individual’s specific voice. These characteristics make identical twins’ voices sound similar.
However, the way in which speakers control their vocal tracts also contributes to generating characteristics such as accent and personality, which can be potentially identified by voice biometrics systems.
This was true for all the technologies used as part of the research, with the Ontario Tech team concluding that voice spoofing between identical twins is not possible when the phrases used are exclusively specific to the candidates, meaning the imposter would be lying. Many set password phrases would not require an imposter twin to lie, given their shared backgrounds such as ‘my home city is’.
In fact, even when repeating exactly the same sentences as the ‘real’ twin, the ‘impostor’ twin only achieved comparison scores between 60 and 75 percent, and only in two sentences. For phrases involving lying such as ‘my favorite color is,’ the impostor score dropped to a minimum.
According to the researchers, the results were such because “voice is just not mere words, it carries the emotional characteristics of the person at the moment of recording.”
Voice tone and the difference in the psychological intent of telling the actual truth and trying to impersonate also reportedly caused a substantial difference in the voice samples.
The researchers were only able to acquire the Software Development Kits (SDKs) of VoiceVantage, so the majority of the tests in the study were conducted using TCH’s algorithms.
For context, having SDKs of a given biometrics system would allow the researchers to run tests using different phrases than the ones pre-recorded and set by the companies (e.g. ‘my voice is my password’).
To circumvent this issue, for Verint and Microsoft Azure, Ontario Tech University chose the closest matching phrases available on their demo platforms.
Speech and voice recognition market set to reach $22B by 2026
New data by ReportLinker suggests the speech and voice recognition market will grow from U.S. $8.3 billion in 2021 to $22.0 billion by 2026, presenting a compound annual growth rate (CAGR) of 21.6 percent.
According to the report, the most impactful driving factors behind the growth would be an increase in the use of smart appliances and the use of artificial intelligence technology to boost the accuracy of speech and voice recognition systems.
In addition, mobile device authentication and control of wearable devices using voice biometrics are also expected to drive the speech and voice recognition market.
In terms of companies leading this shift, Alphabet, Apple, and Microsoft are at the top of the list in the U.S.
The Asia-Pacific region is also set to grow in the forecast period, with many governments eyeing voice biometrics solutions in relation to digitization initiatives.