Telus Digital, the global technology and digital services arm responsible for the telecommunication giant’s call centres, has recently deployed an “accent masking” artificial intelligence tool to change the way its offshore agents sound. The technology analyzes agents’ pronunciation in real time and reshapes their accents to more closely resemble generalized North American or British English speech patterns.
Linguistic profiling or accentism — stereotyping, treating someone unfairly or viewing them negatively based on their accent — can permeate aspects of our society in ways that have real-life consequences.
They can affect hiring decisions, outcomes of legal proceedings, assumptions about the crimes one might commit, assessments in education and access to housing.
Telus appears to be taking advantage of incredible technological advances in signal processing and resynthesis to reproduce an ugly aspect of human behaviour.
As language scholars we believe it is demeaning, manipulative and wrong. And, if we swap in nearly any other social characteristic, it would be discrimination. In order to see racism, we need to see race. So imagine a filter on glasses that changes people’s skin tone, for example, homogenizing our ethnic and racial identities.
There are several strands of evidence that indicate the accentism Telus is practising is not in the best interest of the call centre agents or Canadian consumers.
Some accents are harder
Our discriminatory tastes in speech start early. Children prefer playmates who share their accent, even in linguistically rich cities like Toronto. The credibility and trustworthiness adults attribute to individuals and voice-AI assistants vary with the accent.
Telus says it’s implementing an accent manipulation AI tool because some Telus customers have expressed difficulty in understanding “heavy foreign accents.” This is a leap in logic.
Researchers have known for decades that the perception of accent strength is not well-correlated with how well that voice can be understood. In the case of call centres in India or the Philippines, call agents may be first-language speakers of English, albeit a different variety than Canadian English.
As of yet, there is no public indication that this tech is being used on agents who speak English with a French Québécois accent or a Newfoundland one or a Cape Breton lilt. It’s also worth noting that everyone has an accent; unaccented speech is a myth. An accent is simply a way of speaking that is distinctive to a specific group. If this is genuinely about comprehension, then why would it apply to some accents and not others?
In any instance where we experience difficulty in understanding someone, it is always beneficial to pause and reflect on whether it is a “them problem” or a “me problem.” If there are others who find an accent comprehensible, it is most likely a “me problem.”
(Unsplash)
The cost of accent manipulation
Any human interaction is a two-way street, and the success of that communication comes, in large part, from an implicit collaborative building of understanding — an establishment of a common ground.
When Canadian customers are being duped about who they are talking to — even when this duping is, according to Telus, “to bridge communication gaps and deliver crystal-clear voice experiences” — a cornerstone of that communicative collaboration is removed, leaving an increased opportunity for misunderstanding.
Call centres often already have policies about “regulating identity” of their agents, including strict policies around accents, requirements that agents change their names to something more western-sounding or requirements that agents go through accent modification training.
These requirements can lead to workplace anxiety and stress, affecting the quality of the interaction between the consumer and agent.
Interestingly, call centre agents have already observed that if customers identify the accent as being fake or not genuine — when words and structures being used don’t match the accent — customer relations worsen, and customers become abusive.
What this means is that this “accent softening,” when identified as fake or if leading a customer to assume they are not speaking to a real person, may lead to more frustration by customers and worse treatment of call centre agents.
Exposure to diverse accents
Some might argue that it’s natural to have an easier time understanding someone whose accent is closely aligned with your own. This is true, but this benefit does not come on its own; it is a function of having more experience with an accent.
Think about it this way. There are probably people in your life that you find very easy to understand, like a spouse, family member or close friend. Indeed, that immense familiarity you have with a spouse’s voice makes it both easier to attend to and ignore your loved one. It doesn’t matter if their accent is the same as yours; it’s the experience that matters.
Experience renders both voices and accents more comprehensible. Diverse listening experiences can also make us better listeners, facilitating understanding of a wider range of accents. This is to say, as a listener, you stand to gain from exposing yourself to new voices and accents.
(Unsplash)
AI technology that can modify specific features of an acoustic speech signal while preserving the speaker’s individual identity is cool science. AI can be an incredibly useful tool, but also comes with a human responsibility. Responsible and human-centric approaches to AI should seek to limit harm.
In this case, the “accent softening” of Telus call centre agents is discriminatory to the agents. It’s also a morally dubious misrepresentation of identity to Canadian consumers that disrupts the natural and productive friction that comes with human interaction.
