A collaboration to develop a new approach to telephone speech recognition has been launched in the UK with government backing. The participants in the Advanced Telephone Recogniser project are Marconi Speech & Information Systems, GEC Plc’s Hirst Research Centre and market research company Taylor Nelson AGB. The aim of the project is to speed development of a system capable of automatically generating vocabularies with an almost unlimited number of words. These will be constructed from a single database containing basic sound elements. This technique, known as sub-word modelling, uses algorithms to break down a set of reference words into their component parts. These can then be reassembled to represent different words, forming the required vocabularies. The Hirst Research Centre, part of GEC Research, has already developed a version of sub-word modelling that it says has been used to build off-line whole-word models that can be identified in an ordinary whole-word recogniser. The Advanced Telephone Recogniser project is primarily aimed at developing and expanding this technique. It is planned that a set of extensive trials will be undertaken to gauge its operation over the telephone network. This will be done by building a technology demonstrator with an application chosen from the market research field that makes extensive use of operator intensive telephone communications. Marconi Speech & Information Systems will be drawing on its experience in the development of speech recognition products and the Flexicall Interactive Voice Response system to build the demonstrator and collect new databases from which to extract the sub-word models. Members of the project feel that it will pave the way for greater exploitation of speech recognition technology. This has lost out to other forms of telephone interaction, such as touch-tone keypads, because of the limited speech recognition vocabularies available to date. The ability for users to generate new vocabularies without training is seen as particularly important, since initial applications are likely to be for the market research industry. As the topics of research change, so too will the vocabulary required. The project has been awarded a grant by the UK Department of Trade and Industry.
