TL;DR
This paper investigates what makes a voice "likable" by analyzing acoustic features and subjective ratings. They found that spectral characteristics, particularly those related to timbre, along with pitch and articulation rate, significantly influence likability perception, even more so than prosodic features like intonation. This work lays the groundwork for building predictive models of voice likability.
What Problem Does This Paper Address?
While we intuitively understand that a voice can be perceived as more or less likable, the specific acoustic correlates of this perception are not well-understood. Previous research has often focused on prosodic features like pitch and duration, sometimes neglecting the crucial role of voice timbre. This paper aims to bridge this gap by systematically investigating which acoustic parameters, with a particular emphasis on spectral features related to timbre, contribute to a listener's perception of voice likability. They also sought to validate a simple rating scale for likability and explore the potential for automatic prediction of these subjective judgments.
Methods
The researchers conducted two main studies.
Study 1: Subjective Ratings and Acoustic Analysis * Stimuli: Emotionally neutral read speech sentences from the emoDB database were used to control for content, dialect, mood, and disfluencies. * Rating: Participants rated the likability of speakers on a single seven-point scale ("sympathisch – unsympathisch"). * Acoustic Features: A range of acoustic parameters were extracted using Praat, including: * Spectral features: Center of gravity (S_cog), standard deviation (S_sd), skewness (S_skw), and the 3rd central moment (S_cm3). These are key indicators of timbre. * Prosodic features: Fundamental frequency (f0) and its standard deviation, articulation rate (syllables per second), and intensity. * Analysis: Repeated measures ANCOVA was used to identify significant acoustic correlates of likability ratings. A linear regression model was also employed.
Study 2: Assessing Voice and Speaking Attributes * Purpose: To validate the single likability scale and explore the internal consistency of likability ratings when assessed alongside other voice and speaking attributes. * Stimuli: A subset of stimuli from Study 1 was used. * Rating: Participants completed a comprehensive questionnaire with antonym pairs assessing various impressions of the speaker and their voice/speaking style. * Analysis: Factor analysis was used to identify underlying dimensions of voice perception, and a linear model was used to determine which factors significantly contributed to likability ratings.
Automatic Classification * Purpose: To evaluate the automatic predictability of voice likability. * Features: A large set of acoustic features (988) was extracted using the open-source Emotion and Affect Recognition Toolkit (openEar), including MFCCs, pitch, spectral energy in bands, and formant frequencies. * Classifier: A LogitBoost Alternating Decision Trees (LADT) classifier from the WEKA toolkit was used. * Procedure: A leave-one-speaker-out cross-validation was performed to assess classification accuracy.
Datasets
- emoDB: A database of German emotional speech used for generating the speech stimuli. The authors selected neutral sentences from this database to control for emotional content.
Key Insights and Ideas
- Timbre is King (or at least a major player): The study strongly suggests that spectral features, which are direct correlates of voice timbre, are highly influential in likability perception. This challenges the historical focus on prosodic features alone.
- A Simple Scale Works: A single seven-point likability scale proved to be a reliable and consistent measure for assessing subjective voice perception.
- Gender Matters (but not always as expected): Many acoustic correlates of likability showed significant interaction effects with speaker gender, indicating that what makes a voice likable can differ between men and women. For instance, lower f0 was linked to likability in men, aligning with stereotypes, but a lower spectral center of gravity (darker voice) was perceived positively in both genders.
- Automatic Prediction is Feasible: The automatic classification experiment demonstrated that voice likability can be predicted with an accuracy significantly above chance, highlighting the potential for objective assessment.
Key Results
- Study 1:
- Speakers were rated consistently, and the ranking of speakers was not significantly influenced by listeners.
- Acoustic correlates of likability included:
- Higher articulation rate.
- Lower spectral center of gravity (S_cog).
- Lower f0 (for male speakers).
- Higher spectral standard deviation and skewness (for female speakers).
- A linear regression model explained 45% of the variance in likability ratings.
- Study 2:
- The likability scale showed comparable usage across studies.
- Factor analysis revealed that "warm," "resonant," and "not dark" voice qualities were positively correlated with likability.
- Automatic Classification:
- A leave-one-speaker-out classification achieved 69.66% accuracy and a .769 F-measure for predicting likability.
- The most prominent features for prediction were MFCC means and formant frequencies.
- A classification using the hand-selected features from Study 1 yielded only 62.92% accuracy, suggesting that the automatic feature extraction captured more relevant information.
Limitations and Future Work
- Data Sparsity: The authors acknowledge that gender-separated data analysis in the automatic classification yielded unclear results, likely due to insufficient data. They plan to conduct experiments on larger datasets.
- Stimulus Set Size: The second study used a limited number of stimuli, making it difficult to correlate subjective ratings directly with acoustic parameters.
- Future Work:
- Extend studies to larger datasets, ideally publicly available ones, to build more robust predictive models.
- Further analyze the impact of timbre and speaking style correlates on likability perception.
- Investigate the influence of other acoustic parameters and more sophisticated metrics.
- Explore gender-specific models for likability prediction.
Summary updated