(Not) hearing happiness: Predicting fluctuations in happy mood from acoustic cues using machine learning

Area of Science:

• Affective science
• Computational linguistics
• Machine learning

Background:

• Popular claims suggest virtual assistants may soon detect emotions.
• Prior research shows acoustic features can differentiate emotions and machine learning can detect them.
• Existing work focuses on stable emotion levels, not within-person fluctuations.

Purpose of the Study:

• To investigate if acoustic analysis can automatically detect within-person fluctuations in happy mood.
• To assess the predictive power of acoustic features for real-time emotional states.
• To explore the utility of machine learning algorithms for emotion detection from audio.

Main Methods:

• Collected audio recordings and self-reported happy mood from 20,197 participants across 3 studies.
• Extracted acoustic features from direct speech and ambient sounds.
• Applied neural networks, random forests, and support vector machines for analysis.

Main Results:

• Acoustic features provided minimal predictive insight into happy mood above chance.
• Neither multilevel modeling nor human coders improved mood detection accuracy.
• Machine learning algorithms struggled to identify subtle, everyday emotional shifts.

Conclusions:

• Current acoustic analysis and machine learning methods are insufficient for detecting within-person fluctuations in happy mood.
• Automated emotion detection from voice requires further advancements in affective science.
• Future research should focus on more nuanced acoustic markers for real-time emotion recognition.