Gender Bias in Instruction-Guided Speech Synthesis Models

2025-02-08

AI Summary

This study investigates implicit gender bias in instruction-guided text-to-speech (TTS) models when processing ambiguous occupational style prompts (e.g., “speak like a nurse”). We design a controllable expressive TTS experimental framework integrating prompt-based style interventions, acoustic feature analysis, and cross-model bias comparison. Our work is the first to systematically reveal that mainstream TTS models exhibit stereotyped responses to occupation–gender associations: prompts such as “nurse” or “teacher” consistently trigger significantly higher probabilities of female voice generation, with bias intensity increasing with model scale. Beyond empirically confirming the presence of latent sociocultural bias in instruction-driven speech synthesis, we establish a reproducible, quantitative bias evaluation paradigm grounded in acoustic and behavioral metrics. This provides both theoretical grounding and empirical evidence for developing fair, controllable, and socially aware TTS systems.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Computer Vision: Bias, Fairness & Privacy

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Recent advancements in controllable expressive speech synthesis, especially in text-to-speech (TTS) models, have allowed for the generation of speech with specific styles guided by textual descriptions, known as style prompts. While this development enhances the flexibility and naturalness of synthesized speech, there remains a significant gap in understanding how these models handle vague or abstract style prompts. This study investigates the potential gender bias in how models interpret occupation-related prompts, specifically examining their responses to instructions like"Act like a nurse". We explore whether these models exhibit tendencies to amplify gender stereotypes when interpreting such prompts. Our experimental results reveal the model's tendency to exhibit gender bias for certain occupations. Moreover, models of different sizes show varying degrees of this bias across these occupations.

Problem

• Gender bias in speech synthesis
• Handling vague style prompts
• Amplifying gender stereotypes

Innovation

• Analyzes gender bias in TTS models
• Examines occupation-related style prompts
• Compares bias across different model sizes