Abstract
This paper explores the acoustic and perceptual foundations of "friendliness" in dialogue. By utilizing a multi-dimensional "perception vector" that includes a novel "softness" metric, the researchers quantify the subjectivity of human listeners. A Classification and Regression Tree (CART) model was employed to identify key acoustic drivers, revealing that average pitch (F0 mean) and excitation strength (Ee) are the most significant factors in perceiving friendliness.
Introduction
In real-world interactions, such as customer service, the nuance of "friendliness" often outweighs binary emotional states. However, friendliness is inherently subjective and "fuzzy." This study addresses the lack of a formal framework to quantify this ambiguity and identifies the specific acoustic parameters—pitch, voice quality, and breathiness—that signal friendliness to human listeners.
Methodology: The Perception Vector
To capture the subjectivity of speech, the researchers introduced the Perception Vector. Instead of forcing listeners to choose a single emotion, this vector represents the probability across five basic emotions (anger, happiness, sadness, fear, and neutral) plus a "softness" degree.
A multi-tiered labeling system allowed listeners to provide "first" and "second" choices, which were weighted to calculate the vector. This approach preserves the nuance of human perception, where an utterance might be perceived as both "happy" and "neutral" simultaneously.
Acoustic Analysis and Modeling
The study utilized a corpus of 20 hours of real-world call center recordings. The researchers mapped 28 acoustic features—ranging from prosody (pitch and power) to voice quality (using the Liljencrants-Fant glottal flow model)—to perceived friendliness scores using a Classification and Regression Tree (CART) model.
Key Results and Insights
- Softness is Critical: Friendliness is strongly and positively correlated with "softness"—a property independent of loudness—and "happiness." Conversely, a "very low friendliness" rating typically involves high anger combined with low softness and happiness.
- Predictive Accuracy: The CART model correctly predicted 69% of friendliness categories using 28 leaf nodes.
- Feature Importance:
- F0 Mean (Average Pitch): Identified as the most influential factor in emotion perception.
- Ee (Excitation Strength): The most critical parameter for modeling voice quality.
- F0 Maximum Position: The most important "stress" feature for perceiving friendliness.
Conclusion
By moving beyond discrete emotion labels to a probabilistic perception vector, this research provides a more robust way to model social qualities in speech. These findings offer a roadmap for developing more natural and "friendly" speech synthesis systems by adjusting specific acoustic parameters like pitch and excitation strength.
Summary updated