Spoken Language Translator (SLT) is a project whose long-term goal is the construction of practically useful systems capable of translating human speech from one language into another. The current SLT prototype, described in detail in this report, is capable of speech-to-speech translation between English and Swedish in either direction within the domain of airline flight inquiries, using a vocabulary of about 1500 words. Translation from English and Swedish into French is also possible, with slightly poorer performance.

A good English-language speech recognizer existed before the start of the project, and has since been improved in several ways. During the project, we have constructed a Swedish-language recognizer, arguably the best system of its kind so far built. This has involved among other things collection of a large amount of Swedish training data. The recognizer is essentially domain-independent, but has been tuned to give high performance in the air travel inquiry domain.

The main version of the Swedish recognizer is trained on the Stockholm dialect of Swedish, and achieves near-real-time performance with a word error rate of about 7%. Techniques developed partly under this project make it possible to port the recognizer to other Swedish dialects using only modest quantities of training data.

On the language-processing side, we had at the start of the project a substantial domain-independent language-processing system for English, a preliminary Swedish version, and a sketchy set of rules to permit English to Swedish translation. We now have good versions of the language-processing system for English, Swedish and French, and fair to good support for translation in five of the six possible language- pairs. Translation is carried out using a novel robust architecture developed under the project. In essence, this translates as much of the input utterance as possible using a sophisticated grammar-based method, and then employs a much simpler set of word- to-word translation rules to fill in the gaps.

The language-processing modules are all generic in nature, are based on large, linguistically motivated grammars, and can fairly easily be tuned to give good performance in new domains. Much of the work involved in the domain adaptation process can be carried out by non-experts using tools developed under the project.

Formal comparisons are problematic, in view of the different domains and languages used and the lack of accepted evaluation criteria. None the less, the evidence at our disposal suggests that the current SLT prototype is no worse than the German Verbmobil demonstrator, in spite of a difference in project budget of more than an order of magnitude.