Voice biometrics and speech recognition expand amid countermeasures, bad news for Walmart
Voice biometrics, voice recognition and speech analysis are coming on leaps and bounds, heading to break through the $20 billion barrier as a sector in the next few years. Google looks to be gearing up for further personalization of speech recognition across Android devices while researchers discuss the risks of voice and speech data collection and some of the tools to stop it. Walmart’s bid to have another biometrics lawsuit in Illinois dismissed, this time for voice rather than palm biometrics, as a judge ruled that the privacy claims are plausibly alleged.
Google could be planning further personalization of speech recognition
A teardown by 9to5Google of the Google Assistant app it has uploaded to its Play Store reveals coding that suggests upcoming personalized speech recognition for Google Assistant users.
9to5Google decompiled the Android app file (APK) and found in the Google Assistant settings the description: “Store audio recordings on this device to help Google Assistant get better at recognizing what you say. Audio stays on this device and can be deleted any time by turning off personalized speech recognition. Learn more.”
As 9to5Google explains, such notes and coding are suggestions of upcoming features, and do not necessarily manifest themselves later.
Recent notes from Google explain that it uses federated learning to improve hotword recognition for activation, already with voice files being stored on devices. This next step would go beyond improving the ‘Hey Google’ hotword, to personalized recognition of a user’s commands to the Assistant and particularly with contact names via analysis of recorded interaction, notes the report.
9to5Google states that smart home devices such as newer models of the Nest Hub are already using a machine learning chip to locally process common queries and this approach may now be coming to Android generally.
Privacy would come by opting out, which would delete any recordings.
How voice privacy is being left behind by recognition and replication tech
Citing a report that finds the voice biometrics and speech recognition technology market could reach US$21 billion by 2026, Wired talks to the researchers finding ways to protect people from voice privacy issues.
Voice, choice of words and how a person says them, all reveal significant insights into that person beyond current mood such as any medical conditions and even what they look like, linking to the Speech2Face paper.
A research scientist warns that the information gleaned from voice and speech could be used to build a profile a person which could lead to targeted advertising. Hacking to find a person’s location and voice files would allow for cloning attacks.
Call centers are analyzing callers’ emotions with AI, TikTok collects voiceprints and faceprints of users. Data collection tools are increasingly powerful, but protections are not. And so researchers are developing methods such as voice obfuscation to hide who the speaker is. Beyond simple hardware transformers, speech-to-text-to-speech processes can transcribe dialogue and then repeat it in a different voice.
Distributed and federated learning, rather than centralized pooling of voice data, offers some protection to individuals while the overall system improves. Wired finds that speech and voice anonymization is the most common approach, in two forms.
One tackles language by deleting or replacing “sensitive words” in files. The other anonymizes the voice itself by changing parameters in the sound file such as pitch or swapping in segments from another voice.
However, while this may fool humans, it will not be enough to shake off machine listeners. Only full voice replacement would achieve this, according to researchers. They warn that far more research is going into recognition and replication technologies than protection.
Judge says Walmart must face Illinois law over voice biometrics
In January 2021 Walmart Inc. faced a $10 million fine to around 22,000 current and former Illinois store employees for scanning their palm biometrics and storing the biometric data without meeting the written consent requirement of the Illinois Biometric Information Privacy Act (BIPA).
In July 2021, the same retailer was sued in the same state this time by a worker who claimed headsets used for directing warehouse picking captures and uses voiceprints, without workers’ informed consent, violating BIPA.
This week, U.S. District Judge John Kness said he would not dismiss the proposed class suit against Walmart because the question “Can the retailer’s headset software identify individuals,” reports Law360 (subscription required), is crucial and in the judge’s words “a factual question that’s better addressed after discovery.”
Judge Kness accepts that Walmart may be correct in that the system used, the Honeywell Vocollect Solutions, cannot support the class suit’s privacy claims as it does not collect or use voiceprints for identification, but that the allegations are plausible.
Andrew Barton, the employee who brought the case, amended his complaint to claim that the tech used by Walmart can identify him and his workers. Walmart said the identification comes from employees entering their staff number into the system.
Barton is suing for $1,000 for every negligent BIPA violation and $5,000 for every willful violation found, reports Law360.
In less contentious biometrics use for the retailer, Walmart partnered with Clear to deliver digital COVID-19 vaccination certificates in Clear’s biometric Health Pass.
biometric data | biometric identifiers | biometrics | BIPA | consumer electronics | data privacy | data protection | Google | speech recognition | voice biometrics