This article orginally appeared in the Jul/Aug 1991 issue of Language Industry Monitor “There have been good tone models for British English and Dutch,“ says computational linguist Léon Adriaens, “but there hasn’t been a satisfactory one for German.” At the Institute for Perception Research (IPO) in Eindhoven, Adriaens has been studying German speech for the past seven years. His work culminated at the Eindhoven University of Technology this spring in a doctoral dissertation which introduces an intonation model for the German language. Being able to generate natural-sounding intonation is the key to producing synthesized speech. Intonation is the difference between the tin-voices of science fiction and the increasingly more listener-friendly speech synthesis systems being developed today. To obtain his tone model for German, Adriaens sampled around four hundred sentences taken from about twenty speakers on German radio and television. From this linguistic raw material, Adriaens culled a rendering of the tone model of the language, consisting of an inventory of the relevant pitch movements. He says he accomplished this on the basis of IPO’s methodology of “stylization.”“Using the technique of stylization, we reduce the prosody of natural speech to its essential contours, ignoring arbitrary deviations,” says Adriaens. Researchers at IPO have been working with this technique of stylization since the mid-1960s. They started with Dutch, then moved on to British English, and today Adriaens’ work has given them German. An IPO researcher has also done work on Russian intonation. In Adriaens’ system, the German sentence “Gib mir bitte die Butter” is reduced using the technique of stylization to five straight lines. This approach takes into account behavioral aspects related to pitch — the parameter which carries expressive power. The straight line contours sound the same as the natural pitch curve. “We find people don’t notice the absence the other arbitrary deviations and we can produce a very effective system,” explains Adriaens. Adriaens’ research has been incorporated into the speech synthesis end of the SPICOS system. SPICOS (Siemens Philips IPO COntinous Speech) was a five-year, ESPRIT-driven German language speech processing project carried out jointly by Siemens research labs in Munich, Philips research labs in Hamburg and Brussels, and the IPO in Eindhoven. In the SPICOS prototype, one SUN 3/160 workstation handles analog-digital conversion and spectral analysis for input as well as digital-analog conversion for output. Another SUN 4/260 handling recognition duties, using a predictive language model which operates at phoneme and word level. SPICOS’s Dialogue Handler, running on a Texas Instruments Explorer machine, is the third element. The Dialogue Handler forms the deductive brains of the outfit, being responsible for discourse analysis and answer generation. SPICOS, say the designers, can resolve anaphors, even flag presuppositional failure (Eg, user: “did Jones receive my letter on machine translation?” SPICOS: “There are no letters on machine translation.”) As pre-competitive research, the SPICOS project will not be made commercially available as such. Rather, the knowledge and techniques acquired through the project will percolate through the various labs which participated in it and surface in as yet-unseen products or other projects. Will we be hearing the results of Adriaens’ research in the near future? Possibly. Philips Kommunikations Industrie in Nurenberg is studying the feasibility of a text-to-speech system for German Rail based on his tone model for reading typed messages. Institute for Perception Research, PO Box 513, 5600 MB Eindhoven, the Netherlands; +31 40 75 66 05 |