Intelleca Voice & Mobile, SA`s leading Speech company, has made a massive breakthrough for South African speech recognition with the release of the first South African English acoustic engine built from first principles.
The development is a world first, in that the new acoustic model is the first to be developed locally using local speech science skills, built in cooperation with Scansoft using their world-leading SpeechWorks tool set.
"Intelleca has raised the bar for South African speech recognition with the development, from first principles, of the only industrial-strength acoustic model designed for South African English," says Mike Renzon, Intelleca Voice & Mobile MD.
"The new acoustic engine will ultimately reduce the base error rate of currently used UK engines for SA based speech recognition applications by between 60% and 70%, a level that can support the broad rollout of commercial natural speech applications. The first release of our SA English Acoustic Model is commercially available. It has already reduced the error rate by 30% when compared against the error rate of a UK English engine against SA test speech data. Several releases are planned during the year which will take us to the 70% level and more languages are to follow."
Intelleca has successfully implemented Speech systems for informational type services and basic transactions over the past 3 years for SA companies including Discovery Health, Airports Company South Africa (ACSA), and Property24 a joint venture between Absa and Naspers. "Based on these successes our customers have plans for far-reaching speech applications during 2004 that will require the underlying robustness and accuracy levels of this locally developed engine, " says Renzon. "These applications include advanced transactional services where the user can express dates, times, account information, and amounts using natural speech, directory services, and intelligent call routing where the caller uses natural speech to direct a call to the appropriate place in a company".
According to Dr Etienne Barnard, director of Speech Sciences at Intelleca, "UK and South African English may sound similar, but there are substantial differences that prevent the recognition of even simple words spoken with a South African accent. A model not built on a foundation of local speech can therefore not perform well enough to be successfully used in certain business voice applications. Intelleca`s engine will provide the accuracy and quality to make sophisticated voice applications a feasible solution in call centres and other service industries in South Africa."
The engine is the culmination of two years of work in conjunction with US-based ScanSoft (formerly SpeechWorks). Intelleca`s speech technicians, under the leadership of Dr Barnard, collected several million utterances from their commercial deployments (phrases recorded by South Africans) and manually transcribed and integrated the voice patterns into the engine. "It was a long and labour-intensive job," adds Renzon, "but the result is an acoustic model that delivers greater accuracy and allows us to integrate natural speech recognition into complex applications, such as a telephone directory or call routing."
The same cognitive quality will be built into the Afrikaans and Zulu engines, the next phases in the expansion of local voice recognition technology.
Specialised tasks such as date recognition and deciphering continuous numbers (such as account numbers and credit card details) can now be accurately interpreted by speech technology, opening the door to new industries where speech applications can be reliably applied as front-line customer interaction mechanisms.
In addition to the best voice recognition in the industry, Intelleca`s new acoustic model improves the performance of applications while decreasing hardware demands, allowing customers to extract more power from servers on which the voice technology is run.
The first recipients of the new technology will be Intelleca`s current speech application customers, who will be upgraded to the new acoustic model as soon as possible.
Intelleca and ScanSoft have also joined forces with the CSIR to develop further South African acoustic models for other local languages. Marelie Davel, human language technologies manager at CSIR icomtek, says the group effort will deliver similar models for Afrikaans, Zulu and Sotho. "We hope to have Afrikaans complete by the end of 2004 and work will start on Zulu or Sotho in 2005. Other languages will be incorporated later."
Share