TLDR: A study investigated gender bias in Bark, a Speech-LLM, by analyzing its default speaker assignments for text prompts. Using datasets of gender-stereotyped professions and gender-colored words, researchers found that while Bark is gender-aware with names, it does not exhibit strong systematic gender bias in its speaker assignments. However, it does show some gender inclinations, including counter-stereotypical ones, and infers gender information from text. The study highlights speaker selection as a direct tool for bias investigation in Speech-LLMs.
Large Language Models (LLMs) have become a cornerstone of modern artificial intelligence, demonstrating remarkable abilities in understanding and generating human-like text. However, these models often reflect and even amplify societal biases, including gender bias, which has been a significant area of research. While text-based LLMs encode gendered associations implicitly, Speech-LLMs, which generate spoken language, face a unique challenge: they must produce a voice that inherently carries gendered associations, even when the input text is ambiguous. This makes the process of speaker selection in Speech-LLMs a direct and explicit lens for examining potential biases.
A recent study, titled “Who Gets the Mic? Investigating Gender Bias in the Speaker Assignment of a Speech-LLM,” delves into this very issue. The researchers, Dariia Puhach, Amir H. Payberah, and ´Eva Sz´ekely, focused on Bark, a popular Text-to-Speech (TTS) model based on Speech-LLM technology. Bark is known for its ability to generate expressive and diverse speech, and it automatically assigns a speaker voice when no specific prompt is provided. The core question of the study was whether Bark’s default speaker assignments systematically align with gendered associations, potentially revealing biases embedded in its training data or design.
To investigate this, the researchers developed a novel methodology leveraging speaker assignment as an analytical tool. They constructed two primary datasets: the Professions dataset and the Gender-Colored Words (GCW) dataset. The Professions dataset included sentences featuring stereotypically male and female occupations, adapted from existing bias datasets. The GCW dataset comprised sentences with words that carry gender connotations, ranging from strongly gendered terms like “bloke” to more subtly associated words like “tutu.” Each sentence from these datasets was input into Bark multiple times to account for randomness in its output.
The audio outputs generated by Bark were then analyzed using a Speaker Gender Recognition (SGR) model, which classified the voices as either male or female. To establish a baseline for comparison, two additional tests were conducted. The first, “Professions with Names,” assessed Bark’s basic gender awareness by incorporating gender-conforming names into sentences (e.g., “My name is David. I work as a developer…”). The second, “Neutral Texts,” used standard phonetic passages and Wikipedia abstracts to observe the distribution of male and female voices in a neutral context.
The study also explored Bark’s internal workings by testing how speaker prompts influence gender assignment when bypassing or passing through different layers of the model’s architecture. Bark consists of three main layers: text-to-semantic, semantic-to-coarse, and coarse-to-fine. The text-to-semantic layer is particularly interesting as it encodes speaker prompts and potentially infers gender cues from the text input itself.
The results of the study indicate that Bark does not exhibit a strong systematic gender bias. For the Professions dataset, Bark demonstrated diversity in speaker assignments for the majority of professions, meaning it did not consistently assign a male voice to male-stereotyped jobs or a female voice to female-stereotyped jobs. While some professions did show inclinations (e.g., “mechanic” often assigned a male voice, and “hairdresser” a female voice), interestingly, Bark also produced counter-stereotypical assignments, such as assigning female speakers to professions like “CEO,” “cook,” and “analyst.”
Similarly, for the Gender-Colored Words dataset, about half of the words resulted in neutral speaker assignments. Among the words that did show a preference, many aligned with stereotypes, but some also went against them. The researchers noted that while words like “housewife” or “guy” inherently carry gendered meanings, the systematic association of a female speaker with words like “seashell,” “uptown,” “corset,” “prissy,” and “fussy” could indicate a bias, especially given the negative connotations of the latter two. Despite these specific inclinations, the overall conclusion was that Bark’s diverse speaker assignments for most words suggest it does not have a strong systematic bias.
The study confirmed that Bark is indeed gender-aware when names are explicitly provided in the text, consistently assigning voices that align with the gender of the name. Furthermore, the experiments with speaker prompts revealed that Bark infers gender information directly from the text, specifically within its text-to-semantic layer. This layer plays a significant role in determining the gender of the assigned speaker, and its influence can even diminish the effect of a contradictory speaker prompt.
While the findings suggest Bark does not exhibit strong systematic gender bias under the tested conditions, the researchers emphasize the importance of continued vigilance. The autonomous assignment of speaker gender by Speech-LLMs serves as a powerful diagnostic tool, capable of uncovering subtle patterns in training data that might otherwise go unnoticed. The study also acknowledges limitations, such as Bark and the SGR model operating within a binary gender framework, and the exclusive focus on English despite Bark’s multilingual capabilities.
Also Read:
- AI Agents Explore Human-Like Rhythms: Hormones and Emotions Shape AI Performance and Reveal Biases
- How AI Shapes User Personas: An Analysis of Prompting Strategies in Research
This research provides valuable insights into how Speech-LLMs handle gender in their outputs. By directly observing speaker selection, the study offers a unique perspective on bias investigation in generative AI models. For more details, you can refer to the full research paper: Who Gets the Mic? Investigating Gender Bias in the Speaker Assignment of a Speech-LLM.
