TL;DR
This paper investigates the acoustic properties of overlapping laughter in conversations, finding that it's more prosodically marked than non-overlapping laughter. This effect intensifies with more participants, suggesting entrainment, and people tend to join in laughter with a consistent delay, crucial for dialogue system design.
What Problem Does This Paper Address?
Laughter is a fundamental social lubricant in human conversation, often occurring jointly, leading to overlapping laughter. While speech overlap has been extensively studied, the acoustic characteristics of overlapping laughter and its social implications remain less understood. This paper aims to bridge this gap by exploring how the acoustics of laughter change when it occurs simultaneously with another person's laugh, and how factors like group size influence this phenomenon. Understanding these dynamics is crucial for developing more natural and responsive conversational AI systems.
Methods
The researchers analyzed laughter events from four different corpora: ICSI Meeting Recorder, AMI, HCRC Map Task, and Diapix Lucid. They defined overlapping laughter (OL) as any laugh token that partially or fully coincided with another speaker's laugh token, based on existing annotations. Acoustic measurements were extracted for both overlapping (OL) and non-overlapping (NOL) laughs, including duration, mean and maximum intensity, mean fundamental frequency (F0), and the number of voiced frames. These measurements were normalized using z-scores for each speaker to ensure comparability. The study also analyzed the "delay time" – the interval between the start of one laugh and the start of a subsequent overlapping laugh – to understand response patterns.
Datasets
The study utilized four existing corpora:
- ICSI Meeting Recorder Corpus: Real-life meetings of researchers.
- AMI Corpus: Role-playing and real-life meetings.
- HCRC Map Task Corpus: Task-based dyadic conversations (route finding).
- Diapix Lucid Corpus: Task-based dyadic conversations (spot-the-difference).
These corpora were chosen to provide a diverse range of conversational settings (multiparty meetings vs. dyadic tasks) and to increase the generalizability of the findings. Laughter annotations were used as provided, without re-labeling.
Key Insights and Ideas
- Laughter is More Overlapping Than Speech: Laughter exhibits a higher proportion of overlap compared to speech, underscoring its inherently social and connective function.
- Overlapping Laughter is "Louder" and More Expressive: Overlapping laughs are acoustically more prominent, characterized by longer duration, higher intensity (both mean and maximum), higher pitch (mean F0), and more voiced segments. This suggests a heightened prosodic marking.
- Entrainment in Laughter: The intensification of acoustic features in overlapping laughter increases with the number of participants involved. This points towards a phenomenon of "entrainment" or mutual adaptation between laughers, where they reinforce each other's vocalizations.
- Laughter is Contagious and Group-Dependent: The likelihood of laughter overlap increases with group size, suggesting that laughter is indeed "contagious." However, the "infection" rarely spreads to encompass all participants in a meeting.
- Predictable Laughter Response Delay: People tend to join in overlapping laughter with a consistent delay of approximately 500 milliseconds. This is a critical factor for designing dialogue systems that aim to respond to user laughter.
Key Results
- Acoustic Differences: Overlapping laughs consistently showed statistically significant differences (p < 0.005) in duration, number of voiced frames, and maximum intensity compared to non-overlapping laughs across most corpora. Mean F0 was also significantly higher for OLs. Mean intensity differences were significant for most corpora, though not all.
- Group Size Impact: In multiparty conversations (ICSI corpus), larger group sizes correlated with a higher amount of overlapping laughter time and a greater proportion of participants involved in overlapping laughs. Acoustic measurements also showed that laughs become more prosodically marked as more people laugh simultaneously.
- Delay Time: The average "delay time" for overlapping laughs in dyadic conversations was found to be around 0.489 s (HCRC) and 0.532 s (Diapix Lucid), averaging to approximately 500 ms.
Limitations and Future Work
- Lombard Effect: While higher intensity in overlapping laughs is observed, the authors acknowledge that more research is needed to definitively distinguish whether this is due to entrainment or the Lombard effect (vocal changes in noisy environments).
- Quantifying "Infectiousness": Future work should focus on quantifying the "infectious" nature of laughter to demonstrate that it's driven by contagion rather than just increased overlap probability in larger groups.
- Delay Time and Interpretation: The paper notes that the relationship between variations in "delay times," phonetic properties of laughter, and how these variations affect interpretation needs further investigation.
- Laughter Annotation Quality: The authors mention potential drawbacks of the laughter annotations used, which could impact the analysis.
Summary updated