Talking computers. Spanish work in speech processing

Reliable speech recognition is one of the challenges of user-friendly computing. An Esprit project has brought speech processing in several languages much closer to commercial reality. The authors describe the Polyglot project in which the system for recognising isolated speech in Polyglot builds on a technology originally developed for Italian, extending it to cope with English, French, German, Spanish and Greek. The method consists of three modules: a preselection stage, a fine phonetic analysis and a linguistic module. For each word, the preselection module conducts a fast search across the known vocabulary and selects 2% of the words as the most probable candidates. Fine phonetic analysis hones the possibilities to the best five. Finally, the linguistic module produces a single ´recognised´ word after due consideration of its grammatical category (e.g. whether it´s a noun or a verb), and the category of the preceding word. All this work is done on a PC (386 or higher) with a specially designed DSP (digital signal processing) board. The results obtained varied with different languages, depending on the intrinsic difficulty of the language and the amount of development work undertaken. The authors discuss the Spanish achievements in Polyglot