11 April 2004

South Africa: Computers Learn to Understand Sefrican

Johannesburg — Scientists develop software to recognise local languages - and accents

Thanks to South African boffins, computers have been taught to understand the many languages and accents used in South Africa.

The voice-recognition system, which will one day enable South Africans to speak to machines for routine tasks such as banking and booking flights and hotels, can converse in Xhosa, English (with a range of local accents) and Afrikaans.

"Essentially, we're trying to emulate what happens in the human brain," said Professor Justus Roux, director of the Research Unit for Experimental Phonology at the University of Stellenbosch.

"The system tries to predict what the next element will be; once it's identified a sequence, it goes to a phonetic dictionary."

It took Roux's team of linguists, computer scientists and electrical engineers from universities across the country four years to record thousands of South Africans on the telephone and transcribe the calls phonetically into the computer system.

A white person may say "Durban" while a black South African may pronounce the city more like "Dehban", said Roux.

Voice-recognition systems are being used by the US military - for example for English-Arabic translation handsets in Afghanistan - and commercially in Europe, said Roux, but efforts to adapt them to South African accents had not yet succeeded.

"What we learnt is that there is no short way to gather the data. You must allow for all different pronunciations.

"Age, gender and the tempo - how fast or slow you say words - all have an influence, so your system must allow for lots of variations, not just accents."

The project, which has made the short list of a prestigious new National Research Foundation funding programme, also captured Southern Sotho and Zulu speakers for future use.

In some of the tests, the system's responses sounded so natural that test subjects tried to converse with it, Roux said, but there were still many improvements to be made.

"There are so many nuances in human speech that you cannot capture it all. Our prototype system did break down and did not understand people at times," he said. "But as databases and the technology increase, it will gradually get better."

The next step is for the team to convert the speech-recognition system into a translation system.

Future applications could include commercial use - for instance, company call centres or booking offices - health information hotlines and government services.

The research dovetails with the government's aim of building an electronic gateway that would allow the public to find official information and forms - for example, on births, deaths and marriages - online, Roux said.

A speech-recognition system would allow illiterate people or those without Internet access to get the information on the telephone, he said.

The project was also in line with the government's aims to bring innovation to all South Africans - especially to those in rural areas, said Andrew Aphane, spokesman for the Arts, Culture, Science and Technology Department, a funder of the project.

Voice recognition was the future and was already used in Germany to book train tickets, said Dr Daniel Mashao, an electrical engineer at the University of Cape Town who is working on a speaker-recognition system for security applications.

"One day people will talk to their VCRs and fridges. The immediate benefits are for the disabled community," Mashao said.

"It has even more value in South Africa as it can help us preserve African languages.

"Technology is neutral but it could overrun other languages if it forces people to interact in English," he said.

But Professor Mohlomi Moleleki, chairman of the Pan South African Language Board, had reservations.

"I understand it will play a very important role in multilingualism," he said.

"But if such a system is not managed properly it could become an end in itself and deter people from learning each other's languages."

While building the prototype system, the researchers discovered that Xhosa has more sounds than English or Afrikaans and tends to be spoken in longer sentences.

Even North-West University linguist Professor Daan Wissing, who oversaw the transcription of black English, white Afrikaans and coloured Afrikaans, made new discoveries.

"We found quite a few surprising varieties of pronunciation," said Wissing.

"It was painstaking work but it was fascinating to have two diverse groups - engineers and linguists - working together and to see how the engineers depended on the linguists."

Copyright © 2004 Sunday Times. All rights reserved. Distributed by AllAfrica Global Media (allAfrica.com). To contact the copyright holder directly for corrections — or for permission to republish or make other authorized use of this material, click here.

AllAfrica publishes around 2,000 reports a day from more than 130 news organizations and over 200 other institutions and individuals, representing a diversity of positions on every topic. We publish news and views ranging from vigorous opponents of governments to government publications and spokespersons. Publishers named above each report are responsible for their own content, which AllAfrica does not have the legal right to edit or correct.

Articles and commentaries that identify allAfrica.com as the publisher are produced or commissioned by AllAfrica. To address comments or complaints, please Contact us.