The DictionaryMaker's user interface.

Innovation by digital@SERA's Human Language Technologies (HLT) research group at the
Meraka Institute continues to contribute to the betterment of all South Africans - this time through the newly developed DictionaryMaker. digital@SERA is an IT research platform across the CSIR's Meraka Institute and its Defence, Peace, Safety & Security business unit, and the University of Pretoria's Faculty of Engineering, the Built Environment and Information Technology.

The HLT group has released a system that can be used to create electronic pronunciation dictionaries in any language. The tool guides a speaker who is fluent in a target language, through the dictionary-creation process that can be completed in a fraction of the time it usually takes to create an electronic pronunciation dictionary.

In future, this will be used for speech-based systems to enable ordinary citizens to access information from government services over the telephone. "Our goal is to find a way of making these services accessible to people, as many government services can be provided over the telephone," says HLT research group manager at the CSIR, Dr Marelie Davel.

Pronunciation dictionaries form an integral part of speech recognition or text-to-speech systems. They map the written form of a word to how it should be pronounced. HLT researchers use the sounds of a language, known as phonemes, to create the dictionaries. A phoneme is the smallest phonetic unit in a language that is capable of conveying a distinction in meaning.

The application is the first of its kind in South Africa. "In the past, people used to write words and pronunciations to create dictionaries. The automatic resource we have created, uses an interactive process known as bootstrapping". Bootstrapping describes a process that is self-initiating or self-sustaining. The DictionaryMaker allows a speaker of the target language to "train" the system, which "learns" how words in a language are pronounced, using machine learning techniques (a form of artificial intelligence).

The group has completed systems in isiZulu, Sepedi and Afrikaans, and is currently working on pronunciations for isiXhosa and Setswana. Apart from English, all the official languages in South Africa are known as resource-scarce languages. For these languages, limited electronic language resources exist. The HLT group plans to create pronunciations, speech recognition systems and text-to-speech systems for all the official languages of South Africa within the next two years.

She says the entire system has been released as open source software (OSS) and has already been downloaded by more than a hundred users world-wide. "We have just released it to the international OS community," says Davel. "The science that was necessary to create the system was challenging and we know that this will find application in dealing with real-world needs", she adds.

Davel is confident that the newly developed technology is going to help two groups of people: researchers and engineers who are building speech technology systems and ordinary South Africans who will benefit from speech-based interfaces to information. "For South Africans, this will mean having the option in future of using your language of choice when interacting with speech-based information systems", she says.

Source: CSIR