Investigation of the relationship between turn-taking and prosodic features in spontaneous dialogue
Investigation of the relationship between turn-taking and prosodic features in spontaneous dialogue
复制标题
自发对话中轮流与韵律特征关系的研究
DOI：
10.21437/interspeech.2005-32
复制
发表时间：
2005
影响因子：
0.7
通讯作者：
A. Ichikawa
中科院分区：
文献类型：
--
作者：
Tomoko Ohsuga;M. Nishida;Y. Horiuchi;A. Ichikawa
摘要本研究旨在探讨话轮转换与韵律的关系。我们认为，在实时交流中，说话人要想顺利地进行交流，就必须在话轮边缘之前将话轮转换的前置信号作为韵律特征。我们试图通过决策树方法来区分话轮变化，只使用话轮结尾重音序列中的韵律特征，这些重音序列包括与话轮结尾重音序列相比更早的位置。在辨别实验中，我们使用日语自然对话语料，定义了F0曲线、功率曲线和时长等韵律参数。我们比较了使用有和没有转弯的最终莫拉的参数的两个参数条件。从结果来看，在不使用最终莫拉参数的情况下，其准确率为80%，与使用全部参数时的结果83%相比，并没有显著的下降。考虑到只使用韵律，我们认为这个结果是合理的好。1.在实时通信中，我们可以使用语音进行非常流畅的交互。人们越来越认识到韵律在人与人以及人机交流中的重要作用[1，2]。韵律具有使听者实时、轻松地理解和流畅地控制对话的功能。韵律信息的有效利用使我们对语音理解、语音合成和口语对话系统等技术的发展有了更高的期望。在不同的研究领域中，已经有许多关于话轮转换和韵律信息的研究。例如，句子边界的语调模式与情态和语篇功能有关[3，4，5，6，7，8]。从另一个角度来看，对于实际应用，如人机对话系统，韵律特征用于检测合适的时间forturn-taking或backchannel [9，10，11，12]。大多数研究都表明，在话轮结束时，词汇、句法和韵律信息的特定组合可以作为说话人想保持话语权或想结束话轮的信号，然而，为了判断是否可以进行话轮，听话人不一定要感知到说话人的话语的最后一个音素。我们观察到，话轮转换进行得非常顺利，在连续的话轮之间的延迟最小。在某些情况下，确实存在过搭接短的成功转弯过渡，称为“闭锁”。因此，当考虑到这些现象时，除了在话轮边缘的上述线索之外，我们认为存在更多的话轮转换全局线索。说话者可能会在话轮边缘之前显示出作为韵律特征的前信号，从而使听者清楚地知道说话者是否想在可能的转折点之前更早地完成说话话轮。此外，在我们之前的研究中，我们评估了韵律特征在较早位置对估计句法结构的有效性[13，14]。因此，在本研究中，我们的目标是更积极地对待韵律功能与话轮转换。我们认为，韵律信息可能具有更强的表达说话人对话轮转换的态度的潜力，从这个角度出发，我们尝试只用韵律特征来判断话轮转换是否发生。我们重点研究了既包含话轮结尾莫拉又包含话轮结尾前几个位置的话语的结尾短语，以及在使用除结尾莫拉以外的所有特征的条件下的重复判别。我们使用F0的轮廓、高度和峰值以及功率、持续时间和语速作为韵律参数。
Abstract In this study, we investigated the relationship between turn-taking and prosody. We considered that to interact smoothlyin real-time communication, speakers must show presignalsto turn-taking as prosodic features before turn edges. We at-tempted to discriminate the turn change by the decision treemethod using only prosodic features in turn-final accentualphrases that include earlier positions compared with turn-finalmora. In the discrimination experiment, we used the corpus ofJapanese spontaneous dialogue, and defined prosodic parame-ters such as F0 contour, power contour and duration. We com-pared the two parameter conditions for using parameters withand without the final mora of turns. From the results, the ac-curacy under the conditions of not using the parameters of thefinal mora is 80%, which is not significantly worse than the re-sult of 83% when using all parameters. Taking into accountonly prosody was used, we consider this result to be reasonablygood. 1. Introduction In real-time communication, we can interact very smoothly us-ing speech. There has been a growing appreciation of the im-portant role of prosody in human-human, and also in human-machine communication [1, 2]. Prosody has functions that en-able listeners to achieve real-time and easy understanding, andto control dialogue smoothly. Making effective use of prosodicinformation leads us to the expectations of improvements inthetechnologiesofspeechunderstanding,speechsynthesis,andspoken dialogue systems.In this study, we focus on the dialogue management func-tions of prosody with respect to turn-taking. There have beenmany previous studies on turn-taking and prosodic informa-tion, in various research fields. For example, intonation pat-terns at sentence boundaries are relevant to modality and dis-course functions [3, 4, 5, 6, 7, 8]. From another point of view,for practical applications such as human-machine dialogue sys-tems, prosodic features are used to detect suitable timing forturn-taking or backchannel [9, 10, 11, 12]. Most of these stud-ieshaveshownthatparticularcombinationsoflexical,syntacticand prosodic information in turn-final can function as cues forsignalling that a speaker wants to keep the floor or wants to endthe turn.In order to judge whether it is possible to take the turn ornot, however, the hearer does not necessarily have to perceivethe speaker’s utterance to the last phoneme. We observed thatturn-taking proceeded very smoothly with minimal delay be-tween consecutive speaking turns. In some cases, there werealsosuccessfulturntransitionswithshortoverlap,called“latch-ing”. Therefore, when taking into account these phenomena, inaddition to the above cues at the edges of turns, it is consideredthat more global cues to turn-taking exist. The speakers mightshow presignals as prosodic features before the turn edge, sothat the listeners clearlyknow whether aspeaker wantsto finisha speaking turn at an earlier time before a possible transitionpoint. Moreover, in our previous study, we evaluated the effec-tiveness of prosodic features at earlier positions for estimatingthe syntactic structure [13, 14].Thus, in the present study, we aim to treat prosodic func-tions more positively with respect to turn-taking. We considerthat prosodic information might have some potential to morestrongly express a speaker’s attitude regarding turn-taking.From this point of view, we attempt to judge whether the turnchanged or not using only prosodic features. We focused onthe final phrases of utterances which include not only turn-finalmora but also earlier positions of turn-final, and also attempteddiscrimination under the condition of using all features exceptthe final mora. We used the contours, heights and peaks of F0and the power, duration and speaking rate as the prosodic pa-rameters.