Novauris Unveils New Speech Technologies

Novauris, a company founded a year ago by Dr. James Baker to develop a new generation of speech recognition technology, has unveiled its first demonstration - a system enabling rapid access to any of a quarter of a billion items, equivalent to all addresses in the US.

When a complete name and address is spoken to this system, which runs on a standard PC, it is validated against the 245 million possibilities. The correctly spelled name and address is retrieved together with any associated information (customer account history, etc.). Retrieval typically takes a fraction of a second, and more than 99 percent of the names and addresses are produced fully correctly.

No speaker enrollment is needed.

Former head of BT's speech technology research department, Denis Johnston, who recently tried the Novauris name-and-address system, says: "This is very impressive. There are lots of applications where it would be useful. But it runs deeper, for this level of performance may permanently change how application designers approach dialogue design. Quite simply, it makes speech recognition systems far more attractive to users."

James Baker, who founded Dragon Systems, explains: "I set up Novauris to realize my dream of a new generation of speech recognition technology that can make a significant practical difference to the world. To make that difference, Novauris is eager to work with other speech technology developers and large-scale users of speech technology through licensing, joint development and consultancy."

Baker is celebrated for pioneering the use of hidden Markov modelling (HMM) in speech recognition, now the basis of almost all practical speech recognition systems. While CEO of Dragon Systems, his company introduced and led the market in general-purpose automatic dictation products. His new venture aims to make advances in speech recognition technology of similar magnitude, exploiting a combination of speech knowledge and new statistical techniques.

Most of the Novauris staff are experienced researchers located in Cheltenham, England. They largely constitute the former Dragon Systems UK R&D team, which developed noise-robust speech recognition technology, purchased by Visteon (Ford's former Systems Division), licensed by SpeechWorks and fitted by Jaguar in its cars.

Long-time associates at Dragon and elsewhere, John Bridle and Melvyn Hunt, both internationally respected speech researchers, jointly lead the UK team. "We chose to tackle large-scale name-and-address recognition because we knew there was a demand for it and it was beyond the existing state of the art. Current speech recognition systems are limited to accessing a few million items and need the user to provide a sequence of spoken inputs for each retrieval, so our single-utterance access to over 245 million items represents a major advance. It's also faster, less expensive and more reliable than a human operator in transcribing names and addresses," remarks Dr Hunt.

Partner Bridle adds: "It's also more convenient for the caller, who is spared the need to spell out unusual names." He notes that the technology in the Novauris demonstration, for which patent protection is being sought, can be adapted to many other applications involving extremely large databases.