Bias is one thing, but accent inaccuracy is another.
AI in Crime Prevention
Sustainable Development Tech News
Looking for this week’s news summary?
This week includes free AI training to close the digital divide, EU’s tech startup winning formula, can existing technologies solve the climate challenge and more. Scroll down👇
This week I spotted an article looking at the use of AI to monitor the phone calls of the incarcerated.
The idea is to use the Natural Language Processing (NLP) type of technology we’ve all become familiar with through the likes to Siri or Alexa to record, transcribe and raise alerts based on the contents of recorded conversations between inmates and those they speak with on the “outside”.
These kinds of technologies have boomed in recent years, and when we think of the current hype around the power of AI today, much of this sentiment is driven by the growth in those types of application (although image recognition has come very far too).
The thing is, NLP gets it wrong. A lot.
The article goes on to talk about how prisoners are often subject to technology trials, how machine-interpreted information shouldn’t become part of the judiciary system and how AI driven NLP is significantly worse for Black speakers, than White.
The researchers found that Amazon’s automatic speech recognition software had an error rate for Black speakers that was nearly twice as high as for white speakers.
I’ve been involved in a number of voice-recognition projects, recently in the form of chatbots and a longer time ago in the form of voice controlled cars, and I have to say, “twice as high” is a serious reason for concern.
Search this site for “alexa” and you’ll find a number of articles that talk about my frustration with cutting edge smart speaker technologies. But my pain-affair with “voice-to-text” goes back to last century.
Whilst working for what would now be called an “incubator” in one of the world’s largest telecom operators, I was part of a team that was invited to test drive the latest luxury car with built in voice control. To help set the scene, 3G still wasn’t commonplace at this time, and onboard computers weren’t as powerful as they are today. This meant the voice recognition was rudimentary and error prone.
I made it a few miles in the car before giving up and returning back to base. “Lights on” would sometimes work, but sometimes enable the rear-wiper. “Tune to Radio 1” worked most of the time, so did “wind up windows”. But I’d had enough, it was distracting from the otherwise amazing driving experience.
Next in the car was a friend from Bolton, UK. If you’re not familiar with this accent, check this page out for an example. It’s not the most difficult to understand, however….my friend didn’t make it out of the car park before throwing the keys back with a frustrated “bluddy car!”
It turns out, pretty much every command he issued resulted in the trunk of the car being opened and him having to get out to shut it.
Things have progressed since then, of course, but there are still daily frustrations that I am sure many of us are familiar with.
Now, bringing this back to the above article, if there can be so many issues with the vanilla accents that these models were trained on at the outset, error rates of twice those are surely reasons for serious concern – especially when it comes to decisions of freedom and justice.
Another tool in the belt
Realistically, these tools are used all around us, every day. Zoom calls can now be transcribed on the fly, historic meetings can be analysed to help adjust and hone a sales process, and our written words are routinely analysed for sentiment.
Indeed, I use the voice-to-text service from Otter.ai to record and transcribe notes and meetings, I use Siri and Alexa to control things with my voice, and I even wrote a programme to analyse the sentiment of the messages from my Twitter followers (after realising some of them were spouting nonsense that I’d rather not be associated with).
With all these, there are margins for error. Through experimentation (and frustration) I have discovered these margins and try to work within them.
As a particular example, a couple of years ago I used Siri to dictate a message to my Mum which was meant to say “Do you know her number so I can call her” but actually ended up saying “At doctors, got cancer”. Learning from the trauma that followed, I now try to always double check my messages before hitting send!
I can only hope that implementation of technologies like this are used well within their safety margins, and used (at least with today’s technologies) as another tool to help humans get the job done. Not the only tool.
Sustainable Development Tech News Follows:
GOOD? SHARE IT.
If you found this useful, please do share it and use the form below to sign up for more content like this. Do you have Tech With Purpose or Tech For Good news to share, let me know here.
DO SOME MORE GOOD.
When you use the form below, I will sponsor the planting a tree on your behalf.