Who Finds This Voice Attractive? A Large-Scale Experiment Using In-the-Wild Data (2407.04270v1)

Abstract: This paper introduces CocoNut-Humoresque, an open-source large-scale speech likability corpus that includes speech segments and their per-listener likability scores. Evaluating voice likability is essential to designing preferable voices for speech systems, such as dialogue or announcement systems. In this study, we let 885 listeners rate 1800 speech segments of a wide range of speakers regarding their likability. When constructing the corpus, we also collected the multiple speaker attributes: genders, ages, and favorite YouTube videos. Therefore, the corpus enables the large-scale statistical analysis of voice likability regarding both speaker and listener factors. This paper describes the construction methodology and preliminary data analysis to reveal the gender and age biases in voice likability. In addition, the relationship between the likability and two acoustic features, the fundamental frequencies and the x-vectors of given utterances, is also investigated.

• The paper introduces the CocoNut-Humoresque corpus, a diverse dataset of 1800 YouTube speech segments rated by 885 listeners for in-depth voice likability analysis.
• It reveals significant demographic influences, with male listeners favoring female voices and younger listeners rating voices higher than older counterparts.
• The study shows that while acoustic features like fundamental frequency and x-vectors correlate with likability, they are not sole predictors, informing future voice synthesis design.

An Expert Overview of the "Who Finds This Voice Attractive?" Study

The paper "Who Finds This Voice Attractive?" by Suda et al. presents an in-depth exploration of voice likability using a comprehensive open-source dataset called CocoNut-Humoresque. This dataset is designed to facilitate the paper of voice attractiveness by incorporating subjective likability ratings from a diverse group of listeners who evaluated a substantial number of speech segments. The key contributions of this paper are the construction of the CocoNut-Humoresque corpus and the preliminary analysis revealing biases and tendencies in voice likability.

Construction of the CocoNut-Humoresque Corpus

The CocoNut-Humoresque corpus was created to address gaps in the existing literature on voice attractiveness by including a wide variety of speech segments and listener attributes. Data collection involved 885 listeners who rated 1800 speech segments derived from YouTube content, ensuring a diverse range of voices. Attributes such as gender, age, and favorite YouTube videos were also recorded for both speakers and listeners, allowing for a nuanced examination of likability factors.

The corpus design employs a systematic method for ensuring diversity in listener evaluations. Speech segments were divided into subsets, with each subset rated by at least 11 listeners. An innovative algorithm maximized the diversity of speaker embeddings in each subset to ensure a broad representation of voice qualities.

Analytical Findings and Observations

Gender and Age Biases

A significant portion of the analysis focuses on the gender and age biases in voice likability:

• Gender Biases: It was observed that male listeners tend to give higher likability scores to female voices compared to male voices. Conversely, female listeners exhibited no significant preference based on the speaker's gender but tended to give lower scores overall and demonstrated higher variability in their ratings.
• Age Biases: Younger listeners, particularly those under 30, rated voices higher compared to older listeners. This trend was consistent across both genders, with younger males giving notably higher scores and older females giving lower scores.

These findings underscore the complexity of voice likability, showing that listener demographics significantly influence perceived attractiveness.

Acoustic Feature Analysis

Further analysis explored the relationship between likability and the acoustic features of the voices, particularly fundamental frequencies ($F_0$) and x-vectors:

• Fundamental Frequencies: While some correlation was noted between lower $F_0$ in male voices and higher likability scores from female listeners, this trend did not fully align with expectations from prior studies. Hence, $F_0$ is not the sole determinant of attractiveness.
• X-vectors: The paper utilized $t$-SNE to visualize the x-vectors of speech segments, revealing that likability and opinion variance could be mapped within the x-vector space. This indicates potential for using x-vectors to predict voice likability and variability in listener responses.

Theoretical and Practical Implications

The contributions of this paper have several implications:

• Theoretical Implications: The findings contribute to the understanding of voice perception by highlighting how demographic and acoustic features interact to influence voice likability. This nuanced understanding could guide future research in speech synthesis and voice conversion, aiming to optimize voice qualities for specific listener demographics.
• Practical Implications: From a practical standpoint, the findings can inform the design of more user-preferred synthetic voices in applications such as virtual assistants and public announcement systems. By tailoring voices based on demographic preferences, developers can enhance user experience and engagement.

Future Directions

As the field of AI and speech synthesis evolves, the paper opens several avenues for future exploration:

• Expanded Demographic Analysis: Future work could include a broader range of demographic variables, such as cultural background and linguistic diversity, to provide a more global perspective on voice likability.
• Advanced Acoustic Feature Integration: Incorporating more sophisticated acoustic features and deep learning models could further refine the ability to predict and enhance voice likability.
• Application-Specific Voice Design: Practical implementations of these findings could lead to the development of customizable voice systems that adapt in real-time to user preferences, significantly improving user interaction and satisfaction.

In conclusion, the paper by Suda et al. offers valuable insights into the factors influencing voice attractiveness and provides a robust foundation for both theoretical advancements and practical applications in the field of speech technology. The CocoNut-Humoresque dataset, with its large-scale and detailed annotations, represents a significant resource for ongoing research in voice design and likability analysis.