Classification of Prosodic Events using Quantized Contour Modeling

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 721–724, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Classification of Prosodic Events using Quantized Contour Modeling Andrew Rosenberg Department of Computer Science Queens College CUNY, New York, USA andrew@cs.qc.cuny.edu Abstract We present Quantized Contour Modeling (QCM), a Bayesian approach to the classification of acoustic contours. We evaluate the performance of this tech- nique in the classification of prosodic events. We find that, on BURNC, this technique can success- fully classify pitch accents with 63.99% accuracy (.4481 CER), and phrase ending tones with 72.91% accuracy. 1 Introduction Intonation can significantly vary the intended meaning of a spoken utterance. In Standard American English, contrast is frequently indicated with an accent that has a steeper pitch rise – “I went to the store (not the li- brary)” – than an accent that is used to indicate focus or introduce new information – “I went to the store (before going home)” . At phrase boundaries, rising pitch can indicate uncertainty or that the speaker is asking a ques- tion – “John likes Mary?” vs. “John likes Mary”. Auto- matically detecting prosodic events and classifying their type allows natural language understanding systems ac- cess to intonational information that would unavailable if processing transcribed text alone. The ToBI standard of intonation (Silverman et al., 1992) describes intonational contours as a sequence of High and Low tones associated with two types of prosodic events – pitch accents and phrase boundaries. The tones describe an inventory of types of prosodic events. In this work, we present Quantized Contour Mod- eling, a novel approach to the automatic classification of prosodic event types. In Section 2, we describe related work on this task. We describe Quantized Contour Modeling in Section 3. Our materials are described in Section 4. Experimental results are presented and discussed in Section 5. We conclude and describe future directions for this work in Section 6. 2 Related Work Five types of pitch accents – pitch movements that corre- spond to perceived prominence of an associated word – are defined in the ToBI standard(Silverman et al., 1992): H*, L*, L+H*, L*+H, H+!H*. In addition to these five, high tones (H) can be produced in a compressed pitch range indicated by (!H). For the purposes of the experi- ments described in this paper, we collapse high (H) and downstepped High (!H) tones into a single class leav- ing five accent types. The ToBI standard describes two levels of phrasing, intermediate phrases and intonational phrases which are comprised of one or more intermedi- ate phrases. Each intermediate phrase has an associated phrase accent describing the pitch movement between the ultimate pitch accent and the phrase boundary. Phrase accents can have High (H-), downstepped High (!H-) or low (L-) tones. Intonational phrase boundaries have an additional boundary tone, to describe a final pitch move- ment. These can be high (H%) or low (L%). Intona- tional phrases have five possible phrase ending …

Content truncated to 500 words