Abstract
This study investigates the acoustic-prosodic and lexical characteristics of humor within the context of the sitcom FRIENDS. By analyzing speaker turns labeled via laugh tracks, the researchers identify distinct vocal signatures for humorous speech, including higher pitch peaks and increased tempo. Using these features, machine learning models achieved humor recognition accuracy significantly above baseline levels, demonstrating that the "performance" of a joke is a critical component of its automated detection.
Introduction
While computational recognition of emotions like anger or frustration has seen significant progress, humor remains a complex challenge for social intelligence. Most existing research focuses on textual analysis of puns or one-liners. However, in conversational settings, how a statement is delivered—its rhythm, stress, and intonation—is often as vital as the words themselves. This paper addresses the gap in spoken humor recognition by exploring whether consistent acoustic-prosodic cues can signal humor and be used to build automated classifiers.
Methods
The researchers employed a multi-stage methodology to analyze and classify humor:
- Automatic Labeling: To avoid manual annotation bias, the team used the show's laugh track as a proxy. Speaker turns followed by artificial laughter were labeled "humorous," while others were "non-humorous."
- Feature Extraction: Two primary feature sets were extracted:
- Acoustic-Prosodic: Using Wavesurfer, measurements were taken for pitch (F0), energy (RMS), and temporal features (duration, tempo, and internal silence).
- Lexical: A "bag-of-words" approach was used to count word frequencies, alongside turn length.
- Statistical Analysis: 2-way ANOVA tests were conducted to determine if humor-related prosody varied by speaker gender or individual identity.
- Classification: The ADTree (Alternating Decision Tree) algorithm was used for classification, testing various combinations of lexical, prosodic, and speaker-identity features through 10-fold cross-validation.
Datasets
The study utilized the FRIENDS Corpus, a dataset derived from 75 scenes across 6 episodes of the first two seasons of the sitcom. The corpus consists of approximately 2 hours of audio, totaling 1,629 speaker turns (714 humorous and 915 non-humorous). The data includes the six main cast members and a "GUEST" category for minor characters.
Key Insights: The Sound of Humor
The analysis revealed that humor possesses a distinct prosodic profile. Humorous utterances are generally longer, feature higher maximum pitch, exhibit a wider pitch range, and are delivered at a faster tempo compared to non-humorous turns.
Furthermore, the "sound" of humor appears to be gendered. The differences in mean pitch and pitch range between funny and serious lines were more pronounced in male speakers than in female speakers. The study also found that speaker identity is a crucial factor; accounting for specific vocal habits, such as Chandler’s sarcastic inflections, significantly improved model performance.
Results and Classification
The baseline accuracy for the dataset, achieved by guessing the majority class ("non-humorous"), was 56.2%. The best-performing model reached 64% accuracy (an 8% absolute improvement) by combining lexical, prosodic, and speaker features.
Notably, models using only 13 prosodic features performed nearly as well (~60% accuracy) as those using over 2,000 lexical features (~61-63%). This underscores the importance of vocal performance in humor. The classifier was also more effective at identifying humor in male speakers (14% improvement over baseline) than in female speakers (5% improvement).
Conclusion and Future Work
While the study successfully identifies acoustic signatures of humor, it is limited by its source material. Sitcom actors may use "hyper-articulation" to ensure jokes land, potentially making prosodic cues more obvious than in natural speech. Additionally, the laugh track proxy may miss visual or subtle humor. Future research should explore more sophisticated NLP techniques for irony detection and test these findings across different media genres and cultures to determine the universality of these "humor signatures."
Summary updated