Interactive Spoken Language Education – Annual Report 1999

In the second year of project ISLE, work has focussed on several areas:

• collection of a detailed, annotated corpus of non-native speech for training and testing purposed
• extensive testing of the recognition and diagnosis components
• design and implementation of the final ISLE Demonstrator
• complete documentation of the ISLE technology, to facilitate easy integration with other packages

In the final three months, the project will perform a validation of the demonstration system with a population of language-learners, in order assess the utility of the ISLE software, the intuitiveness of the demonstration interface, and the quality of the recognition and diagnosis.

In its second year, the ISLE project has performed extensive testing of the speech recognition and error-diagnosis components with a corpus of non-native speech that closely models what real Italian and German learners of English would produce. This testing has shown that performance depends highly on the level of the speaker; students with very poor English may experience difficulty in being recognized, while those with especially good English may find that very few speech errors are detected. Nonetheless, this leaves a considerable number of potential users in the intermediate-level range. As determined in Project Year 1, there is a large market for "edutainment" software within the European Union as well as a demand for tools to assist in second-language learning, especially when that language is English. The ISLE software is well positioned to help fill this void in the market.

The core of the ISLE technology is provided by Entropic Ltd.’s IHAPI speech recognition toolkit and by custom-build ISLE diagnosis routines for finding and explaining pronunciation errors. IHAPI provides robust recognition of native speech, and when used with an initial adaptation, recognition levels for non-native speakers approach those for native speakers. Pronunciation errors are detected based on an expectation strategy: Because many of the errors to which intermediate level students succumb are predictable based on their mother tongue or on the irregularities present in English, a relatively small set of errors can be targeted. Finally, the undeniably important yet difficult to solve problem of stress-based errors are handled by the ISLE stress component, which discovers cases in which the student has placed stress on the incorrect syllable of a word.

A fully-functional example application, showcasing the ISLE technology, has been constructed by project-partner Dida*El, a company with many years experience in producing such multi-media based software. This Demonstrator embeds the ISLE tools in a traditional English-learning package. Students follow the central character through a series of interconnected travels and events, and practice their language skills using sets of speech-based as well as non-speech-based exercises relating to each of the topics. The feedback given to the student in the ISLE Demonstrator is highly friendly, requiring no special knowledge of phonetic symbols, thus allowing the student to concentrate on improving his or her pronunciation skills rather than on interacting with the system.

In the final months of the project, until April 2000, the ISLE project will be tested on a group of English-learners, in order to see the system at use in a real environment, rather than in laboratory conditions. In addition, the components will be fully documented, in order to make the adoption of ISLE technology by other interested parties as easy as possible.