Text-to-speech conversion for German text with TALKMAN

Abstract

Since 1985 the research institute of microprocessor engineering is treating a project, whose primary aim is to develop communication aids for speech impaired people. In the course of this project the problem has been tackled to produce rather high quality speech starting from written text. This paper describes an algorithm which converts plain German text into a detailed speech description, which can be used directly to control conventional phon synthesizers. The conversion takes place on word level. Obviously the key problems of the conversion task are related closely to the specific peculiarities of the language. In the case of German we are confronted with an outstanding morphological complexity. It is very usual in German to construct new words by combining existing ones. So morphological decomposition becomes the most important operation. As a consequence this paper will concentrate on the morphological analysis step of a single German word. The TALKMAN algorithm is based on a dictionary of about 7000 morphs and other word fragments. The principal idea is to search for reconstructions of the input word by elements of the dictionary. Provided the dictionary is near to complete, in general more than one reconstruction will be found. Here some heuristics is involved to identify the most plausible decomposition. In the dictionary additional information concerning intonation and pronunciation is stored for each member. After the word structure has been discovered this information is processed in subsequent steps to produce the phonetic description and the stress pattern of the input word as well. This description contains enough information about speech to compute the parameters which control the phon synthesizer by a straight-forward interpretation.