Page 12 - EngineerIT June July Issue 2026

LARGE LANGUAGE MODELS

LOST IN TRANSLATION:
LOST IN TRANSLATION:

Why AI still struggles with
Why AI still struggles with

South African conversation
South African conversation

By Justin Render, with insights from
Warren Hawkins, Managing Director
of Euphoria Telecom

rtificial intelligence has
become remarkably good
Aat turning speech into
text. With relative accuracy, it
transcribes meetings, records
customer interactions, summarises
conversations and generates
searchable records in seconds.
Still, even as technology improves,
South African speech patterns
can be uniquely difficult for AI to
understand.

"Like most people at present, I use
transcription software every day,"
says Warren Hawkins, Managing
Director of Euphoria Telecom,
a cloud-based telephony and
contact centre solutions provider.

Transcription tools have become
essential to most of us in our daily
workflow. Having gone paperless
myself, I can’t remember the
last time I took notes manually
in a meeting or interview. The "It’s a given that I can talk to an AI, and it gets things wrong. But then my
technology saves time, improves marketing manager could say exactly the same sentence, and it's perfectly
productivity and allows information captured," says Hawkins.
to be captured far more efficiently
than traditional note-taking. But I That inconsistency becomes more pronounced when AI encounters the realities of
am acutely aware of its limitations. South African communication. Accents, regional dialects, indigenous languages
and code-switching all introduce layers of complexity that global models often
struggle to interpret correctly.

The challenge is not unique to South Africa. AI systems have had to learn regional
accents and variations in English all over the world. The difference, Hawkins
argues, is that South African speech patterns are still underrepresented in the
datasets used to train many of today's leading models.

Warren Hawkins, Managing Director of Euphoria Telecom

12 | EngineerIT June/July 2026