How CALL-SLT works

Like many other web applications, CALL-SLT uses a client/server architecture. Nearly all the processing takes place on a remote server, currently in Geneva. The client, which is running in your web-browser, isn't doing much more than recording what you say, sending it to the server, and displaying the results it gets back.

When you speak, the server carries out the following main stages of processing:

1. The first step is to carry out speech recognition with the commercial Nuance Toolkit. We use small-vocabulary speech recognition modules specially constructed for the CALL-SLT task. They are built, using the Regulus platform, on top of grammars that define what words and phrases the recognizer is allowed to consider. The platform makes it easy for us to change the vocabulary, for example if we want to add new words to the system.

2. The recognizer produces several different guesses as to what the student said. The server processes each guess and then decides which one to pick. Its choice depends partly on how “forgiving” you have set recognition to be.

3. Processing of each recognition guess starts with syntactic and semantic analysis. The server then translates the result into a language-neutral form, and compares it with the language-neutral version of the thing you were asked to say.

The examples you hear when you click on “Help” were created by people who have used the system. When users are registered as native speakers, the system automatically records correct matches and saves them for other students to listen to later.

Why are we different from commercial systems?

People sometimes ask us what we offer that's not included in commercial language learning packages like Rosetta Stone. If you've used one of these systems, you'll soon see the difference. Their pronunciation practice tools give you a specific phrase to say at each turn and score you on how well you imitate it. You can't say anything else. With CALL-SLT, you can answer in many ways and experiment with flexible use of language to build up your fluency.

Publications

If you're curious to learn more about how CALL-SLT works, you might want to check out some of our recent papers.