Resources for turn competition in overlap in multi-party conversations: speech rate, pausing and duration

Emina Kurtić, Guy J. Brown, Bill WellsInterspeech 2010

TL;DR

This paper investigates how speakers use speech rate, pausing, and the duration of simultaneous talk to compete for conversational turns in multi-party settings. By analyzing real-world meeting recordings, the researchers found that simply continuing to speak beyond a certain point (around 2 words or 0.5 seconds) is the most significant indicator of turn competition, with pausing also playing a role.

What Problem Does This Paper Address?

Conversations are typically a smooth dance of taking turns, but sometimes this rhythm breaks down, leading to simultaneous speech, or "overlap." While overlap can be non-competitive (like agreeing with someone), it can also be a battleground for control of the conversation. Previous research has focused on fundamental frequency (pitch) and energy (loudness) as cues for this "turn competition." However, the authors argue that other crucial prosodic features – specifically speech rate, pausing, and the duration of the overlap itself – have been underexplored in this context. Understanding these cues is vital for building more natural and intelligent conversational agents and for a deeper understanding of human interaction.

Methods

The researchers employed a Conversation Analysis (CA) approach to categorize overlaps as either competitive or non-competitive. This involved human annotators carefully examining conversational extracts and using established criteria for interactional behavior.

To quantify the prosodic features, they extracted data from the ICSI Meetings Corpus. Key techniques included:

  • Speech Rate (SR) Computation: Measuring syllables per unit of time, comparing in-overlap speech rate to a speaker's normal rate, and analyzing changes in rate before and during overlap.
  • Pausing (PAU) Features: Analyzing the frequency, position, and length of pauses within overlapping speech, and whether overlaps occurred immediately before or after a pause.
  • Duration (DUR) Features: Measuring the length of overlapping talk in both time and word count, and examining the timing of overlap onset relative to the ongoing turn.

These extracted features were then fed into Decision Tree (DT) classifiers. The DTs were built using a "leave-one-in" approach, where each feature group (SR, PAU, DUR) was tested individually, and then a "leave-one-out" approach, where feature groups were removed from a comprehensive model to assess their combined importance. This allowed them to identify which features and combinations were most predictive of competitive overlap.

Datasets

The study utilized a dataset of 1455 overlap instances extracted from eight meetings within the ICSI Meetings Corpus. These meetings involved 6-8 participants each and were spontaneous conversations. The researchers specifically focused on overlaps involving two native American English speakers and excluded backchannels, collaborative, and choral productions, which are typically non-competitive.

Key Insights and Ideas

  • Duration is King: The most powerful predictor of turn competition is simply how long someone continues to speak during an overlap. Overlaps longer than approximately 2 words or 0.5 seconds are highly likely to be competitive. This aligns with the idea that persisting in overlap signals a strong desire to maintain or gain the floor.
  • Pauses Play a Supporting Role: While duration is dominant, pausing also contributes to identifying competitive overlaps. The frequency of pauses within an overlap seems to be a relevant cue.
  • Speech Rate is Less Clear-Cut: Contrary to some previous hypotheses, the study found that overlappers (those initiating the overlap) don't necessarily use faster speech rates to signal competitive intent at the outset. Overlappees (those whose turn is being overlapped) might use speech rate to respond, sometimes increasing it in non-competitive overlaps and decreasing it in competitive ones, but this is a less robust indicator than duration.
  • Redundancy in Features: Many of the analyzed features, particularly within the duration and pausing groups, are correlated. This suggests that while multiple cues might be present, the core signal for competition often boils down to a few key indicators.

Key Results

  • Overall Accuracy: Decision trees trained on all features achieved a classification accuracy of 66.19% for overlappers and 65.36% for overlappees, significantly outperforming a majority baseline of 51.68%.
  • Duration's Dominance: When used alone, duration features achieved accuracies of 66.10% (for overlappers) and 65.95% (for overlappees), demonstrating their strong predictive power. Removing duration features significantly degraded the performance of the overall model.
  • Pausing's Contribution: Pausing features also showed significant improvement when used alone (e.g., 57.88% for overlappers), but their contribution became redundant when combined with duration features.
  • Speech Rate's Limited Role: Speech rate features alone did not yield significant improvements over the baseline for overlappers. For overlappees, while SR features combined showed some improvement, individual SR features were not strong predictors.

Limitations and Future Work

The authors acknowledge that their study focuses on specific prosodic features and that other cues, such as fundamental frequency (F0), intensity, and lexical content, also play a role in turn competition. They plan to investigate the interaction of the features studied here with F0, intensity, and overlap onset position in future work. Additionally, they aim to explore the relationship between these prosodic findings and the lexical and pragmatic properties of overlapping turns. The current study also relies on human annotation for categorization, and future work could explore how these features might be used in automated systems for detecting and managing conversational overlap.

Summary updated

BibTeX
@inproceedings{kurtic2010resources,
  author = {Emina Kurtić and Guy J. Brown and Bill Wells},
  title = {{Resources for turn competition in overlap in multi-party conversations: speech rate, pausing and duration}},
  booktitle = {Interspeech},
  year = {2010},
}

Loading related content...

Figure