AUTOMATIC RECOGNITION OF SENTENCE TYPE FROM PROSODY IN DUTCH Vincent J. van Heuven*, Judith Haan** and Jos J.A. Pacilly* *Phonetics Laboratory, Department of Linguistics and Holland Institute of Generative Linguistics, Leiden University Cleveringaplaats 1, PO Box 9515, 2300 RA Leiden, The Netherlands **Department of General Linguistics and Dialectology, Centre for Language Studies, Nijmegen University Erasmusplein 1, PO Box 9103, 6500 HD Nijmegen, The Netherlands ABSTRACT This paper investigates to what extent statements, Wh- questions, Yes/No-questions and declarative questions in Dutch can be automatically discriminated on the basis of global and local F 0 -parameters. Global para- meters were the slope and mean pitch of upper and lower trend lines that were fitted through F 0 -curves; local parameters were onset and offset F 0 of a termi- nal question-marking pitch rise. Results indicate that women mark the interrogative status of a sentence more often and perceptually more saliently. Generally, global downtrend parameters are better predictors of sentence type than parameters of the final rise. 1. INTRODUCTION Dutch, as well as a host of other languages, distin- guishes at least two sentence types: statements (S) and questions (Q). Questions are further subdivided into three types: W Wh-questions (also called ‘information questions’), beginning with a question word (who, where, when, what, why, etc., and often followed by in- version of subject and finite, Y Yes/no-questions, syntactically marked by inversion of subject and finite, D Declarative questions, which have the same lexical items and word order as the corresponding state- ment. Lexico-syntactic marking of interrogativity, therefore, is strongest in W, weaker in Y, and absent in D. All three Q-types are claimed to be prosodically marked; we expect the phonetic/prosodic interrogativity mar- kers to be stronger in inverse proportion to the (num- ber of) lexico-syntactic markers. Cross-linguistically, question intonation has always been strongly associated with a local terminal rise in pitch. In part of the literature question intonation is still largely identified with this final rise [1,2]. Still it would seem more accurate to say that questions are universally marked by the presence of a ‘high’ ele- ment somewhere in the sentence [3]. This high pitch may manifest itself both locally, e.g. in the initial, medial or final portion of the utterance, and globally, either in the guise of a raised register 1 [4,5,6] or of the absence of F 0 -downtrend; presence of F 0 -down- trend is commonly observed in and across statements [7,8]. The distribution of high elements may differ be- tween and within languages. In Dutch, the Q-shibolet is a terminal rise, which is, in fact, the only Q-marker acknowledged in current models of Dutch intonation [9,10]. Yet, there are early claims in the literature that the Dutch question contour is hammock-shaped, i.e., it has a high beginning, a low stretch in between and an equally high ending [11,12]. It has also been claimed that Dutch questions are realized in a higher register [11,13]. Recent exper- iments have indeed brought to light that Dutch ques- tions differ considerably from statements in terms of pitch range and downtrend [14]: F 0 -level at the onset of the utterance (higher for Q than for S), degree of global F 0 -downtrend (downwards for S vs. more level or even upward for Q), height of F 0 -maxima in prominent syllables (greater in Q than in S), and level of the low pitches in non-prominent syllables (greater in Q than in S). Since most of these Q-markers develop as the utter- ance progresses, it is important to determine whether there is a point in time in the utterance before the onset of the terminal rise where differences are suffi- ciently large to safely identify the sentence type as S, W, Y, or D. The purpose of the present research, then, is to determine to what extent sentence type in Dutch can be automatically recovered from the prosody of the utterance, at some early point in the sentence, i.e., before the presence vs. absence of a sentence-final pitch rise provides a decisive cue. 2. METHOD 2.1. Speech materials. The basic materials are a database of 800 Dutch 1 Register is defined here as an area within a given speaker's overall pitch range, enclosed by the highest and lowest frequency within which tones of a particular utterance are realized [15]. EUROSPEECH ’97 5 th European Conference on Speech Communication and Technology Rhodes, Greece, September 22-25, 1997 ISCA Archive http://www.isca-speech.org/archive