A cross-linguistic study on turn-taking and temporal alignment in verbal interaction

Spyros Kousidis, David Schlangen, Stavros SkopeteasInterspeech 2013

TL;DR

This paper investigates how speakers manage turn-taking and synchronize their speech timing across three linguistically diverse languages. They found that despite grammatical differences, speakers in all languages tend to avoid gaps and overlaps in conversation, and exhibit temporal alignment. This suggests that smooth turn-taking and alignment are universal conversational goals, potentially driven by cultural conventions rather than solely linguistic structure.

What Problem Does This Paper Address?

Conversations are a fundamental aspect of human interaction, and a key feature is how participants take turns speaking. While it's generally understood that people try to avoid awkward silences (gaps) or talking over each other (overlaps), the precise mechanisms and universality of this "turn-taking" are less clear. Furthermore, researchers have observed that speakers often unconsciously synchronize their speech patterns, a phenomenon known as "alignment." However, previous studies on these temporal features (gaps, overlaps, pauses) have often been limited to a few languages or specific methodologies, leaving open questions about cross-linguistic variation and the factors influencing alignment. This paper aims to address this by conducting a cross-linguistic study on turn-taking and temporal alignment in three genetically and regionally diverse languages, exploring whether linguistic properties or cultural factors play a larger role.

Methods

The researchers employed a semi-automatic approach to analyze audio recordings of conversations. First, they used the Praat software to segment speech and silence intervals. This was followed by manual annotation to correct segmentations, assign speech to speakers, and mark non-speech events. From these annotations, timelines were created to visualize turn alternations.

Key temporal features were then extracted:

  • Gaps: Silent intervals between different speakers' utterances.
  • Pauses: Silent intervals within a single speaker's utterance.
  • Overlaps: Intervals where speakers talk simultaneously.
  • Floor Transfer Object (FTO): A combined measure of gaps and overlaps, where overlaps are represented by negative durations.

To assess temporal alignment, they correlated speaker averages of these temporal features within individual interactions. They also explored various predictor variables (like joint active time and pause length) to understand their influence on FTO and to rule out alternative explanations for the observed correlations.

Datasets

The study utilized three distinct corpora of spontaneous spoken interactions, collected under similar experimental conditions:

  • Cabécar: A Chibchan language spoken in Costa Rica.
  • Fongbe: A Gbe language spoken in Benin.
  • Georgian: A Kartvelian language spoken in Georgia.

These languages were chosen for their genetic and geographical diversity, and for their differing typological properties relevant to predicting the end of an utterance (EUP). The corpus consisted of:

  • 19 dialogues (58 min) in Cabécar (7 pairs).
  • 31 dialogues (119 min) in Fongbe (7 pairs).
  • 24 dialogues (60 min) in Georgian (8 pairs).

Key Insights and Ideas

  • Universal Drive for Smooth Transitions: Despite significant grammatical differences between the languages, speakers in all three consistently minimized gaps and overlaps in turn transitions. This suggests a universal preference for smooth conversational flow.
  • Alignment is Pervasive: Interlocutors in all three languages showed significant correlations in their temporal features (like average FTO length), indicating that they synchronize their speech timing.
  • Cultural Conventions Over Linguistic Structure? The study found that linguistic properties (like word order rigidity or prosodic cues for EUP) did not reliably predict the observed differences in turn-taking latencies. For instance, Cabécar, expected to have more predictable turn-taking due to its structure, actually exhibited longer floor transfers. This hints that cultural conventions might play a more significant role in shaping turn-taking behavior than grammatical features alone.
  • Alignment is Not Just About Engagement: While factors like "joint active time" (how much speakers are talking) correlate with FTO and can "compress" it, they don't fully explain the observed temporal alignment between speakers. This suggests alignment is a more nuanced phenomenon than simply being engaged in the conversation.
  • Pause Length is Less Predictive of Alignment: Intra-sentence pause length showed weak correlations with FTO and alignment, suggesting that alignment is more closely tied to the transitions between turns (gaps and overlaps) rather than pauses within turns.

Key Results

  • FTO Distributions: Georgian had the smallest average FTO length (mean -96 ms) with a high overlap percentage (51%). Cabécar had the longest average FTO length (mean 332 ms) with a low overlap percentage (23%). Fongbe had a balanced distribution (mean 91 ms) with a high gap percentage (64%).
  • Inter-speaker Correlations: Average FTO length (AFT) was significantly correlated between speakers in all three languages (r values ranging from 0.53 to 0.69). Average gap length (AGL) and overlap length (OVL) also showed significant correlations in some languages.
  • Predictor Variables: Joint active time (JAT) was strongly correlated with FTO across all languages, indicating that more engaged dialogues lead to compressed FTO. However, other variables like intra-sentence pause length (APL) showed weak predictive power for FTO.

Limitations and Future Work

  • Methodological Limitations: The paper acknowledges that previous methods for quantifying temporal alignment (like "sliding window") have arbitrary choices. While they used novel approaches, the interpretation of temporal alignment phenomena remains an open problem.
  • Cultural Factors: The study suggests cultural conventions are important but doesn't deeply explore how these conventions manifest or are learned.
  • Future Directions: The authors propose further cross-linguistic research to identify the specific components that determine turn-taking latencies. They aim to investigate how grammatical, typological, semantic, and cultural factors interact to influence alignment, moving beyond macroscopically observable correlations to understand the micro-level mechanisms. They also plan to explore the role of non-verbal cues like gaze in this process.

Summary updated

BibTeX
@inproceedings{kousidis2013cross,
  author = {Spyros Kousidis and David Schlangen and Stavros Skopeteas},
  title = {{A cross-linguistic study on turn-taking and temporal alignment in verbal interaction}},
  booktitle = {Interspeech},
  year = {2013},
}
Figure