AI feels your frustration trying to be understood behind a mask. Literally
As almost everybody knows at this point in the pandemic, masks fuzz voices in noticeable and infuriating ways. Now, it is possible for an AI algorithm also to tell the difference (without the complaining, so far).
Seven scientists from four research labs in China working together say they have achieved moderate success in differentiating between someone speaking with and without a mask.
In the same vein, Alcatraz says it has an AI mask detection product that will do the same task visually. More on that below.
The audio endeavor was one of three tasks posed during the 11th annual Computational Paralinguistics Challenge. The motivation behind the mask sub-challenge was to automatically spot who is not covering their potentially coronavirus-tainted exhalations.
The winning team reported that its method achieved 78.8 percent accuracy on the unweighted average recall (UAR) metric. Scientists found that wearing a mask changes the normal muscle constriction, vocal effort and transmission loss associated with speech.
Those results were not particularly exciting, though they were a bit better than a baseline system, according to the researchers’ Arxiv, not-peer-reviewed, paper.
The labs involved were Duke Kunshan University, Wuhan University’s computer science school, AI Lab of Lenovo Research and Sun Yat-sen University. Work was funded in part by China’s National Natural Science Foundation and by regional science and technology programs.
Hours of recorded ordinary speech by 32 native German speakers — the Mask Augsburg Speech Corpus — were broken into one-second clips for training and evaluation. They were artificially corrupted by warping speech speed and randomly erasing audio. The researchers also applied SpecAugment, a technique that masks acoustic features.
Informing the algorithm of individual speakers’ gender increased the AI’s performance, but no conclusions were offered for why.
As for Alcatraz, company executives say their Rock biometric facial authentication system can detect the presence of a mask, an increasingly important capability for businesses and building operators. They say the Rock interoperates with “any access control system.”
Passive three-dimensional sensors and machine learning to spot people “tailgating” at entry points to gain access or avoid detection in some way. In the process of monitoring, the company claims, the software knows someone is wearing a mask before they use their badge to pass through a check point.