LISTENING TO LISTS: STUDYING DURATIONAL PHENOMENA IN ENUMERATIONS Hannes Pirker and Stefan Kramer {hannes,stefan}@ai.univie.ac.at Austrian Research Institute for Artificial Intelligence (ÖFAI) Schotteng. 3, A-1010 Vienna, Austria September 22, 1999 Abstract A study on durational phenomena in list like enumerations in German is presented. Due to its highly structured and uniform nature this rather specialized utterance type seems especially well-suited for investigating principles of the rhythmical or- ganization of speech. A corpus extracted from radio weather reports is used in or- der to investigate phenomena like prefinal lengthening and effects of isochrony and prominence. In addition to studying durational phenomena with standard statisti- cal methods, the data also was analyzed using Structural Regression Trees (SRT), a machine learning algorithm. 1 Introduction The temporal organization of speech has been an issue amongst phoneticians for decades. Numerous phonetic studies dealed with the investigation of (segmental) duration. Typ- ically these works dealed with the isolated influence of single factors on duration and established a whole inventory of possible influences that span from segmental context to linguistic factors (see, e.g., [6] for a summary of factors and [5] for discussion and further references). In spite of this rich phonetic tradition the specification of segmental duration is still an issue in speech synthesis. In this context not only the factors but also their exact quantitative effects have to be determined. Because of the number of factors and their complex interactions this becomes a very difficult task. Nowadays large corpora and sta- tistical approaches are used to tackle this problem. Nevertheless, all these sophisticated methods – including neural networks and machine learning techniques (for German e.g. [5], [13], [14], [10]) – have to struggle with the problem of data sparsity produced by the misbalance between the vast number of possible value combinations and the restricted size of available corpora. Thus, it is necessary to integrate as much phonetic knowledge as possible in order to sustain the automatic methods and to restrict the number of hypotheses generated by these methods [15]. In this spirit the work presented in this paper is an exploratory study in the context of an upcoming large scale study on segmental duration in Austrian German. Besides the 1