What can laughter tell us about meeting effectiveness?

During meetings, there is both verbal and non-verbal communication. Even just a meeting’s audio stream contains a wealth of information beyond the spoken words [1]. Laughter may be the most important example of this non-verbal communication. It is an essential part of the human experience; no culture is known to exist without it [2]. According to one study [3], 8.6% of the time a person vocalises in a meeting, it is because they are laughing.

Why measure laughter?

The presence of laughter can be a sign of well-being and contentment at work. It can indicate engagement and good group cohesion. It may be useful feedback about the enjoyability of a presentation. But why stop there? Consider analysis of laughter trends, as well. Understanding the differences in amounts of laughter within your company’s meetings means understanding team dynamics and highlighting the presence of hierarchical barriers. Trends in an individual’s laughing frequency, on the other hand, may be an indicator of rising stress or imminent burnout.

How to measure it?

Early efforts in laughter detection focused on finding identifiable acoustic signatures. This proved difficult, however, given the highly variable nature of laughter [4]. There are studies that rely on complex apparatuses to detect laughing – one used a combination of a tri-axial accelerometer, a positioning device, and a proximity sensor, for example [5]. Other researchers developed their own wearables for the purpose [6]. However, a less intrusive and more modern solution comes in the form of machine learning models, which try to categorise sound based on more parameters than humans could simultaneously keep in mind. These require an abundance of labelled audio data to recognise and learn such nuances. Fortunately, several suitable datasets are publicly available [7][8].

What does Meeting Canary do?

We are implementing one such model, an adaptation of the deep convolutional network used by Jon Gillick and colleagues [9][10]. The original paper was primarily concerned with extracting the laughter audio, whereas we only need to flag its presence and duration. One of the most immediate benefits to our product will be the ability to separate crosstalk instances into bursts of group laughter versus interruptions. The endeavour of laughter detection is our first foray into having AI work on audio, instead of transcribed text. With this hurdle crossed, there may soon be a lot more we can analyse from the audio directly.

What are the challenges?

Even with the latest models, laughter detection is far from a perfected science yet. For one, none of the publicly available datasets used for training differentiates between all the different types of laughter. Whether sarcastic, awkward, or mocking, to name a few, laughter can convey a whole range of emotions. Furthermore, it is unlikely that one’s use case has complete parity with the conditions of the training data. For instance, the Switchboard [7] dataset contains phone conversations between two persons at a time (albeit thousands of them), so there may be difference for larger group meetings.

Another complication comes in the form of intermixed speech and laughter; 0.8% of a person’s vocalisations in meetings are words spoken while laughing [3], potentially blurring the lines for detection software. Finally, while all cultures have laughter, its significance and acceptability may differ.

In conclusion, while it may sound whimsical at first, laughter is a potential vehicle for all kinds of information that can make your meetings better. And after all, that is what we aim to do for you.

References

1. M. T. Knox & N. Mirghafori. (2007). Automatic Laughter Detection using Neural Networks. Interspeech.
2. H. Tanaka & N. Campbell. (2014). Classification of Social Laughter in Natural Conversational Speech. Computer Speech & Language.
3. K. Laskowski & S. Burger. (2007). Analysis of the Occurrence of Laughter in Meetings.
4. J. Trouvain. (2003). Segmenting Phonetic Units in Laughter. Proceedings of the 15^th International Congress of Phonetic Sciences.
5. H. Hung, G. Englebienne & J. Kools. (2013). Classifying social actions with a single accelerometer. Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing.
6. A. Shimasaki & R. Ueoka. (2017). Laugh Log: E-textile Bellyband Interface for Laugh Logging. Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in Computing Systems.
7. J. J. Godfrey & E. Holliman. (1997). Switchboard-1 Release 2. Linguistic Data Consortium.
8. N. Morgan, D. Baron, J. Edwards, D. Ellis, D. Gelbart, A. Janin, T. Pfau, E. Shriberg & A. Stolcke. (2001). The meeting project at ICSI”. Proceedings of the First International Conference on Human Language Technology Research.
9. K. Ryokai, E. D. López, N. Howell, J. Gillick & D. Bamman. (2018). Capturing, Representing, and Interacting with Laughter. CHI
10. J. Gillick, W. Deng, K. Ryokai & D. Bamman. (2021). Robust Laughter Detection in Noisy Environments. Interspeech.