Accent and dialect discrimination is a pervasive problem in Britain, with the way a person speaks often being erroneously and unjustly tied to their level of intelligence and trustworthiness.
As conversational AI is increasingly adopted across public services, there is a danger of unequal treatment between accents and dialects being reinforced, as systems demonstrate more difficulty in understanding ways of speaking which differ from the "standard".
Dr Chris Montgomery, Senior Lecturer in Dialectology at the University of Sheffield, said that while the cause of linguistic bias in AI is not explicit prejudice, “the results might be the same” as if it were.
He added: “It may so happen that those speakers of varieties that have prejudicial responses from humans might also be the ones that have less representation in training and testing data.
“And so there's a problem that the outcome essentially is to reproduce unequal treatment of voices, even if it's unintentionally due to a training data gap.”
This means that those with less “standard” ways of speaking may have more difficulty accessing the support they need.
However, research is being undertaken as part of a partnership between the University of Sheffield and AI business ICS.AI to help develop a framework which would allow conversational AI systems to better understand linguistic variations.
The project is built upon the socio-linguistic principle that language variations are structured rather than random, and the resulting conclusion that they contains patterns which technology can be taught to recognise. Consequently, researchers are able to investigate which features are more likely to differ, as well as which are most likely to be misrecognised, in the specific service context.
Instead of treating such variation as “noise” in need of being controlled out, the team seeks to map relevant differences, identify where systems struggle, test with real users, measure discrepancies in performance, and build feedback loops to ensure continuous improvement.
Dr Montgomery, who is leading the initiative, said that there are two stages in the pipeline from a person calling a number to receiving a service in which AI bias can surface.
The first relates to how well the system registers certain pronunciations and rhythms of speech in the speech-to-text, or transcription stage. The second concerns variations in the language used itself, in the form of colloquial phrasing or non-standard grammar, and how well the system is able to understand this in the interpretation stage.
By overcoming these obstacles, the team seeks to guarantee “the ability of speakers to access the same quality of service regardless of how they speak”.
This not only means training systems to better understand a wider range of people, but “a graceful handling of uncertainty”. This could take the form of a system asking for clarification, offering alternatives, or transferring to a human adviser faster when it is unable to interpret speech, rather than cutting the call or asking a person to repeat themselves several times.
Going forward, the team hopes to transition from theory to testing, developing evaluation methods to test conversational AI across a range of speakers. They believe that the framework they create could also be used in a number of other European countries, where regional accent and dialect variations are also deeply entrenched.
Dr Montgomery said: “[I’d like that, in] five to 10 years' time, we're no longer really asking whether public service AI copes with accents and dialect variation. It does.
“What I'd hope is that linguistic inclusion becomes part of the sort of procurement process of part of the testing and quality assurance.
“So public bodies, councils are actually asking, has this system been tested by the people that we serve? Does it perform well across different communities? What happens when it gets things wrong?”
ICS.AI’s agents are currently being used by a number of local authorities including Derby City Council, Renfrewshire Council, and Bristol City Council. These live environments are helping the company refine its approach to linguistically inclusive conversational AI, so that it can better understand and respond to natural variations in how residents speak.