Learning to Predict Pitch Accents and Prosodic Boundaries in Dutch

Learning to predict pitch accents and prosodic boundaries in Dutch Erwin Marsi1, Martin Reynaert1, Antal van den Bosch1, Walter Daelemans2, V´eronique Hoste2 1 Tilburg University ILK / Computational Linguistics and AI Tilburg, The Netherlands {e.c.marsi,reynaert, antal.vdnbosch}@uvt.nl 2 University of Antwerp, CNTS Antwerp, Belgium {daelem,hoste}@uia.ua.ac.be Abstract We train a decision tree inducer (CART) and a memory-based classifier (MBL) on predicting prosodic pitch accents and breaks in Dutch text, on the basis of shal- low, easy-to-compute features. We train the algorithms on both tasks individu- ally and on the two tasks simultaneously. The parameters of both algorithms and the selection of features are optimized per task with iterative deepening, an efficient wrapper procedure that uses progressive sampling of training data. Results show a consistent significant advantage of MBL over CART, and also indicate that task combination can be done at the cost of little generalization score loss. Tests on cross-validated data and on held-out data yield F-scores of MBL on accent place- ment of 84 and 87, respectively, and on breaks of 88 and 91, respectively. Accent placement is shown to outperform an in- formed baseline rule; reliably predicting breaks other than those already indicated by intra-sentential punctuation, however, appears to be more challenging. 1 Introduction Any text-to-speech (TTS) system that aims at pro- ducing understandable and natural-sounding out- put needs to have on-board methods for predict- ing prosody. Most systems start with generating a prosodic representation at the linguistic or sym- bolic level, followed by the actual phonetic real- ization in terms of (primarily) pitch, pauses, and segmental durations. The first step involves plac- ing pitch accents and inserting prosodic boundaries at the right locations (and may involve tune choice as well). Pitch accents correspond roughly to pitch movements that lend emphasis to certain words in an utterance. Prosodic breaks are audible interrup- tions in the flow of speech, typically realized by a combination of a pause, a boundary-marking pitch movement, and lengthening of the phrase-final seg- ments. Errors at this level may impede the listener in the correct understanding of the spoken utterance (Cutler et al., 1997). Predicting prosody is known to be a hard problem that is thought to require informa- tion on syntactic boundaries, syntactic and seman- tic relations between constituents, discourse-level knowledge, and phonological well-formedness con- straints (Hirschberg, 1993). However, producing all this information – using full parsing, including es- tablishing semanto-syntactic relations, and full dis- course analysis – is currently infeasible for a real- time system. Resolving this dilemma has been the topic of several studies in pitch accent placement (Hirschberg, 1993; Black, 1995; Pan and McKe- own, 1999; Pan and Hirschberg, 2000; Marsi et al., 2002) and in prosodic boundary placement (Wang and Hirschberg, 1997; Taylor and Black, 1998). The commonly adopted solution is to use shallow infor- mation sources that approximate full syntactic, se- mantic and discourse information, such as the words of the text themselves, their part-of-speech tags, or their information content (in general, or in the text at hand), since words with …

Content truncated to 500 words