<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-3010">
  <Title>Turn-taking in Mandarin Dialogue: Interactions of Tone and Intonation</Title>
  <Section position="2" start_page="0" end_page="72" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Fluent dialogues require effective turn transitions between speakers. Research in turn-taking, typified by (Duncan, 1974), posits several key signals for turn-taking, including a turn-change signal which offers to cede the floor, speaker-state signal which indicates taking the floor, withinturn signal, and continuation signals. This process fundamentally requires cooperation between participants both to produce the contextually appropriate signals and to interpret those of their interlocutor. These analyses have proposed a wide range of cues to turn status, ranging from gaze and gesture in a multi-modal context to prosodic cues including pitch, intensity, and duration as well as lexical and syntactic cues.</Paragraph>
    <Paragraph position="1"> Much of this fundamental research as well as computational implementations have focused on English, a language with well-studied intonational sentence and discourse structure. A substantial body of work has identified sentence-like units as well as fragments and repairs in conversational speech, including (Ostendorf, forthcoming; Liu et al., 2004; Shriberg et al., 2001) These approaches have employed lexical and prosodic cues in diverse frameworks, including Hidden Markov Models employing decision trees and hidden state language models, neural networks, and maximum entropy models. (Shriberg et al., 2001) identified jump-in points and jump-in words in multi-party meeting speech using prosodic and language model features at accuracies of 65 and 77% under equal priors. Furthermore, (Ward and Tsukahara, 2000) demonstrated that backchannels occurred at predictable points  with specific prosodic characteristics in both English and Japanese.</Paragraph>
    <Paragraph position="2"> In the current paper, we consider the interaction of potential prosodic intonational cues related to turn-taking with the realization of lexical tone in a tone language, Putonghua or Mandarin Chinese. Mandarin employs four canonical lexical tones distinguished by pitch height and pitch contour: high level, mid-rising, low falling-rising, and high falling. We explore whether prosodic features are also employed in turn-taking behavior in this language and whether the forms are comparable to those employed in languages with lexical tone. We demonstrate that intonational cues quite similar to those in English are also employed in Chinese with lower pitch and intensity at ends of turn units than at the beginnings of those turn units. Interruptions likewise are distinguished from smooth turn transitions by prosodic means, including greater pitch elevation.</Paragraph>
    <Paragraph position="3"> We demonstrate how these changes interact with lexical tone by substantial lowering of average pitch height across all tones in final positions and contrast pitch contours in final and non-final positions. Finally, these cues in conjunction with silence and durational features can be employed to distinguish turn-unit final words from non-final words in the dialogue and words that initiate interruptions from those which start smoother turn transitions.</Paragraph>
    <Paragraph position="4"> In the remainder of the paper we will briefly describe the data set employed in these experiments and the basic extraction of prosodic features (Section 2). We then present acoustic analyses contrasting turn unit final and turn unit initial syllables under different turn transition types (Section 3). We will describe the impact of these intonational cues on the realization of lexical tone (Section 4). Finally we will apply these prosodic contrasts to enable classification of words for finality and interruption status (Section 5).</Paragraph>
  </Section>
class="xml-element"></Paper>