The prediction of focus

We present results an how focus is marked intonationally in Gennan. Several speakers produced a /arge corpus of sentences. The corpus was constructed in a way that sentence modality and place of focus could only be differentiated by infonational means. Acoustic features representing the infonational parameters pitch, duration, and intensity, were extracted manually or automatically. The relevance of these features and the effect of , several Iransfonnations were tested with statistical methods. Perceptual experiments where the listeners had to judge the naturalness and categories of the utterances were perfonned as weiL By calculating average values for the (appropriately transfonned) relevant features we found "normal'~ prototypical cases. We will show that by Zooking at utterances where all listeners agreed on the naturalness and (intended) categories we arrived at coinciding results. At the same time we found "unusual" but regular productions. MATERIAL AND PROCEDURES This paper is concerned with the prediction of focus; focus is the part of an utterance which is semantically most important. On the phonetic surface focus is marked by the focal accent (FA). To be more exact, we will try to predict the phrase that carries the FA. Our material consists of 360 German utterances, spoken by 6 untrained speakers (3 male, 3 female). Three different sentences with a similar syntactic structure were each put in different contexts that determined sentence modality as weil as place and manner of focus (simple focus, focus projection, or multiple focus). Fora detailed description of the corpus and the intended focal structures see the relevant contributions in /3/. In each of the sentences the last two phrases could be stressed, depending on the surrounding context. Based on the sentence modality system according to Altmann /1/, the sentences formed minimal pairs that could only be differentiated by their intonational form: focus in final vs. focus in prefinal position on the one hand, and questions vs. non-questions on the other hand. Table 1 shows an example of a context sentence, the pertinent test sentence, and the induced sentence modality and place of focus. Table 2 shows the three test sentences, a word-by-word translation into English, an appropriate translation, and a finer description of the induced sentence modalities questionjnon-question (Q/NQ). The only instruction given to the speakers was to produce the context and the test sentence. We did not instruct the speakers to produce the FA in a certain way. By instructing the speakers, one can eliminate certain variabilities and facilitate the analysis. On the other hand one loses the chance to find regular and interesting deviations and merely receives several realizations of representative cases where representativeness is based on the intuition of the researcher. By evaluating a relatively !arge number of cases we expected to find both representative cases (which we will call central types) and rarer but acceptable cases (which we will call marginal types). We evaluated our data in two ways that proved to be converging: Strategy 1: We extracted acoustic feature values that represent the prosodic parameters pitch, duration, and intensity. Using a statistical classifier we tested the relevance of the features with respect to the place of the FA. By calculating average values for the relevant features we found the centrat type of each QjNQ-FA constellation. Elmar Nöth" .. Lehrstuhl für Informatik 5 (Musterekennung), Friedrich-Alexander-Universität,Erlangen,F.R.G. Table 1: Example of context and test sentence, induced sentence modality, and place of focus Constellation of sentence modality and focus: Assertion, focus on "linen" Context: Mother: "What does the master make Nina weave at the moment?" Sentence: Employee: "She makes Nina weave the linen." Table 2: Test sentences, translation, and induced sentence modalities Sie läßt die Nina das Leinen weben ?/. She makes the Nina the linen weave She makes Nina weave the linen assertive question vs. assertion Lassen Sie den Manni die Bohnen schneiden ?/! Make the Manni the beans cut Make Manni cut the beans polar question vs. imperative Lassen wir den Leo die Blumen düngen ?/! let us make the Leo the flowers fertilize let us make Leo fertilize the flowers polar question vs. adhortative Strategy 2: We presented the utterances to a forum of Iistencrs who judged the naturalness, category, and place of FA. Category roughly means sentence modality. As for the differences cf. /3/. By selecting the utterances that were judged to be the "best" ones and by comparing the feature values of those utterances with the average values from strategy 1 we found the central type as weil as marginal types. EXTRACTION OF FEATURES For each utterance we calculated the following features: For the whole utterance The fundamental frequency (F0) at the end of the utterance (offset). The all point regression line of the F0 values (reg). The duration in centiseconds (dur). For the 2nd and 3rd phrase The maximal and minimal F 0 value ( max2, min2, max3, min3 ). The difference of the position an the time axis of the extreme values in centiseconds (pos2, pos3). The duration in centiseconds (dur2, dur3). The average and maximal logarithmic energy (aint2, mint2, aint3, mint3). The parameter values were extracted "by hand" on mingograms and automatically from the digitized versions of the utterances. (See /7/ for details on the Fo.-algorithm and the computation of the energy values.) In /5/ we showed that automatically extracted F0 values produced recognition rates comparable to those from mingogram values. An automatic extraction of the durational values however would pose a problern (see below).