⚡ electrical engineering

Sexualised synthetic personas encode and amplify gendered power asymmetries through voice

Drawing on a Feminist HCI perspective, this study demonstrates that commercial AI voice systems reinforce gendered power asymmetries by encoding female-coded voices with sexualized and submissive traits while associating male-coded voices with dominance and positive attributes.

Original authors: Alice Ross, Ariadna Sanchez, Elin Kanhov, Catherine Lai, Eva Szekely

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alice Ross, Ariadna Sanchez, Elin Kanhov, Catherine Lai, Eva Szekely

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a digital toy store where you can buy a "voice" instead of a teddy bear. You can pick a voice that sounds like a "Southern Gentleman" or a "Parisian Temptress." You can even tell the voice to giggle, sigh, or sound "flirty." This is what a popular company called ElevenLabs offers.

This research paper is like a group of detectives (the researchers) walking into that store, buying a bunch of these voices, and then asking a diverse group of regular people: "What do these voices actually make you feel?"

Here is what they found, broken down simply:

1. The Setup: The "Voice Menu"

The researchers looked at the most popular voices on the site. They noticed the menu was split into two main sections:

The "Serious" Section: Voices labeled "Educational," "News Anchor," or "Tech Expert."
The "Sexy" Section: Voices labeled "Flirty," "Temptress," or "Seductive."

They picked the top male and female voices from both sections. Then, they made a clever switch: they took the "sexy" voices and made them read boring, neutral text (like a science passage). They also took the "serious" voices and made them read the same boring text. This was like testing if a "sexy" voice sounds sexy even when it's talking about math, or if a "serious" voice sounds serious even when the words are neutral.

2. The Experiment: The "Word Game"

They played these audio clips for about 120 people in the US and Canada. The listeners had a big list of 36 words (like charming, creepy, dominant, shy, seductive) and had to pick the three that best described what they heard.

3. The Big Discovery: The "Gender Script"

The results showed that the voices weren't just different; they were following a very old, very strict script.

The Male Voices (The "Boss" Script):
Whether they were reading a sexy script or a boring science script, the male voices were mostly described as dominant, confident, and positive. Even when they were trying to sound "flirty," people still heard a guy in charge.
- Analogy: Think of a male voice like a sturdy oak tree. Whether it's in a storm or a sunny day, it still feels solid, strong, and reliable.
The Female Voices (The "Submissive" Script):
The female voices were almost always described as submissive, shy, and highly sexualized. Even when they were reading the boring science text, people still heard them as "flirty" or "seductive."
- Analogy: Think of a female voice like a rubber duck. No matter what you put it in (a bathtub or a kitchen sink), it's still expected to squeak, float, and be "cute." The researchers found that the female voices had extra "fluff" added to them—like breathy sighs and giggles—that made them sound submissive, regardless of what they were actually saying.

4. The "Content" vs. The "Tone"

The researchers wanted to know: Is it the words that make the voice sound sexy, or is it the sound itself?

For Men: The words mattered a lot. If a male voice read a sexy script, people called him "seductive." If he read a boring science script, people called him "serious." The words changed the perception.
For Women: The words barely mattered. Whether a female voice read a sexy script or a boring science script, people still called her "seductive," "intimate," or "sensual."
- The Takeaway: The "sexiness" was baked into the voice's tone and breathing, not just the words. The female voice was "programmed" to sound like a fantasy, even when it was just reading a textbook.

5. What the Listeners Said (The "Real Talk")

The researchers also asked people to write down their honest thoughts.

About the Men: Some people thought the "sexy" male voices sounded "threatening" or "predatory," while others thought the "serious" ones were just "boring" or "flat."
About the Women: Many people felt the "sexy" female voices were cringe-worthy. They used words like "forced," "trying too hard," and "annoying." One listener said it sounded like "a teenage boy's idea of what sexy sounds like."
- The Metaphor: It was like watching a bad actor trying to play a seductress. It felt like a caricature—an exaggerated, fake version of femininity that didn't feel real or human.

6. The Conclusion: Why This Matters

The paper argues that these AI voices are like digital mannequins. They aren't showing us the full, messy, diverse reality of how real men and women speak. Instead, they are reinforcing old stereotypes:

Men are the bosses (dominant, positive).
Women are the objects (submissive, sexualized).

The researchers warn that even though these AI voices are marketed as "creative tools" for making video games or movies, they are actually copying and amplifying harmful biases. They are taking the "male gaze" (how men imagine women should sound) and turning it into a product that anyone can buy.

In short: The paper found that while AI voice technology could be a tool for freedom and diversity, right now, it's mostly acting like a mirror that only reflects outdated, stereotypical ideas about gender, making women sound like submissive fantasies and men sound like dominant bosses.

Technical Summary: Sexualised Synthetic Personas Encode and Amplify Gendered Power Asymmetries Through Voice

Problem Statement
While digital technologies offer potential spaces for identity exploration and sexual empowerment, online discourse remains pervaded by toxic masculinity, heteronormativity, and the harassment of women and LGBTQ+ individuals. Concurrently, the emergence of commercial Text-to-Speech (TTS) and "Voice AI" platforms has introduced a new arena where human-like voices are generated, personalized, and sold as products. Unlike traditional voice assistants designed to reduce friction in functional tasks, these platforms encourage users to "play Pygmalion," prompting and customizing voices with paralinguistic features (e.g., giggles, sighs) to fit specific, often sexualized, personas.

The core problem addressed is whether these commercial systems, despite offering a seemingly diverse range of prompts (including "flirty," "seductive," and "submissive" for both male and female-coded voices), inadvertently reproduce and amplify existing gendered power asymmetries. Specifically, the authors investigate if the synthesis of "sexualised" voices reinforces binary and heteronormative stereotypes, encoding unequal power dynamics where female voices are associated with submission and male voices with dominance.

Methodology
The study employs a Feminist HCI perspective to analyze a listening experiment involving a diverse group of North American English-speaking listeners.

Stimuli: Audio samples were generated using ElevenLabs' "Voice Library." The researchers selected popular voices categorized under sexualized adjectives (e.g., "flirt," "temptress") and non-sexualized adjectives (e.g., "informative," "presenter"). For each gender (male/female), the most popular voice in each category was selected.
Experimental Design: A 2x2 factorial design was used to isolate the effects of vocal style from linguistic content.
- Voice Style: Sexualized (flirt/temptress) vs. Non-sexualized (informative/presenter).
- Text Content: Sexualized scripts (prompts provided by the platform) vs. Neutral text (excerpts from the modified Rainbow Passage).
- This resulted in 10 distinct voice personas, each recorded with both text types.
Procedure: Participants (N=120, recruited via Prolific) listened to 30 randomized trials. For each audio sample, they selected the three most suitable adjectives from a balanced list of 36 words (categorized as positive, negative, dominant, submissive, and sexualized). They could also provide optional free-text comments.
Participant Demographics: Participants were categorized into four groups based on gender identity and reported attraction to ensure representation of non-heterosexual orientations: (1) Women attracted to women, (2) Women attracted only to men, (3) Men attracted to men, and (4) Men attracted only to women.
Analysis: The study utilized quantitative analysis via generalized linear mixed effects regression models (using lme4 in R) to examine adjective distribution. Qualitative analysis was performed on 280 free-text comments. Acoustic analysis included calculating mean Fundamental Frequency (F0) and speaking rate.

Key Results
The analysis reveals systematic asymmetries in how listeners perceive male- and female-coded voices, even when linguistic content is controlled.

Gendered Adjective Distribution:
- Male-coded voices were significantly more likely to be described with dominant ( $p < .001$ ) and positive ( $p = 0.0027$ ) adjectives.
- Female-coded voices were significantly more likely to be described with submissive ( $p < .001$ ) and sexualized ( $p < .001$ ) terms.
- Negative adjectives were distributed roughly equally across genders.
Impact of Text vs. Voice:
- For male voices, listener perception was highly content-dependent. When sexualized text was used, adjectives like "seductive" or "creepy" were common; when neutral text was used, perceptions shifted to "confident" or "serious."
- For female voices, the perception of being "seductive," "sensual," or "intimate" persisted regardless of whether the text was sexualized or neutral. This suggests that prosodic and paralinguistic features (e.g., breathiness, sighs) inherent in the female-coded "sexualized" voices drive the sexualized perception more than the spoken words.
Listener Demographics:
- Group 4 (men attracted only to women) showed distinct patterns: they were more likely to apply sexualized adjectives to sexualized female voices and less likely to apply negative adjectives to them compared to other groups. They were also more likely to choose positive words for sexualized male voices.
Acoustic Findings:
- Sexualized male voices had a lower mean F0 (70.08 Hz) compared to informative male voices (110.88 Hz).
- Sexualized voices (both genders) were spoken at a slower rate (2.4 syllables/sec) compared to informative voices (3.8 syllables/sec).
Qualitative Feedback:
- Comments on female voices often described them as "annoying," "forced," or "trying" to be sexy, with some participants noting they felt like "male-authored fantasies of femininity."
- Comments on male voices ranged from "bland" to "predatory" or "rapey," though a significant portion criticized them as "monotonous."

Key Contributions

Empirical Evidence of Bias: The study provides quantitative and qualitative evidence that commercial Voice AI systems, even when offering "gendered parity" in prompt categories, encode and amplify binary and heteronormative stereotypes.
Decoupling Content and Style: By varying text and voice independently, the authors demonstrate that the sexualization of female-coded voices is driven significantly by prosodic and paralinguistic features, not just the semantic content of the speech.
Feminist HCI Perspective: The work applies a Feminist HCI lens to TTS, highlighting how technologies are shaped by and encode gender relations, specifically showing how "sexual empowerment" in AI interfaces often defaults to submissive female and dominant male tropes.
Methodological Innovation: The study introduces a novel listening test combining adjective selection with diverse participant demographics (including non-heterosexual groups) to evaluate synthetic voice perception.

Significance and Claims
The authors claim that these findings are significant because they reveal that the "creative" customization offered by Voice AI platforms often reinforces dated, cisnormative, and heteronormative social biases rather than enabling genuine diversity.

Reinforcement of Stereotypes: The "sexualised" personas (e.g., "The Parisian temptress") are identified as auditory caricatures that lack the agency or control found in human sex work, instead presenting a clichéd, heteronormative schema of vocal attractiveness.
Market Implications: As these voices are marketed for creative industries (games, film), they risk supplanting human voice actors while perpetuating narrow ideals of gender.
Societal Impact: The authors argue that as conversational AI becomes normalized, these representations may influence users' own speaking styles through linguistic accommodation and acoustic-prosodic entrainment, potentially normalizing these power asymmetries in human-AI and human-human interactions.
Modesty of Scope: The paper explicitly limits its claims to the specific segment of the market examined (ElevenLabs' Voice Library) and acknowledges that other systems (e.g., xAI's Grok, OpenAI's ChatGPT) exhibit similar issues regarding harassment and exploitation, suggesting a broader industry-wide problem. The authors do not claim to have solved the issue but emphasize the necessity of examining how artificial voices are adopted and perceived in society.