The Binding Effect: Analyzing How Multi-Dimensional Cues Form Gender Bias in Instruction TTS

Kuan-Yu Chen; Yi-Cheng Lin; Po-Chung Hsieh; Huang-Cheng Chou; Chih-Fan Hsu; Jeng-Lin Li; Hung-yi Lee; Jian-Jiun Ding

eess.SPMar 21, 2026

The Binding Effect: Analyzing How Multi-Dimensional Cues Form Gender Bias in Instruction TTS

Authors: Kuan-Yu Chen, Yi-Cheng Lin, Po-Chung Hsieh, Huang-Cheng Chou, Chih-Fan Hsu, Jeng-Lin Li, Hung-yi Lee, Jian-Jiun Ding

Organizations: Graduate Institute of Communication Engineering, National Taiwan University, Taiwan · AI Research Center, Inventec Corporation, Taiwan

Abstract

Current bias evaluations in Instruction Text-to-Speech (ITTS) often rely on univariate testing, overlooking the compositional structure of social cues. In this work, we investigate gender bias by modeling prompts as combinations of Social Status, Career stereotypes, and Persona descriptors. Analyzing open-source ITTS models, we uncover systematic interaction effects where social dimensions modulate one another, creating complex bias patterns missed by univariate baselines. Crucially, our findings indicate that these biases extend beyond surface-level artifacts, demonstrating strong associations with the semantic priors of pre-trained text encoders and the skewed distributions inherent in training data. We further demonstrate that generic diversity prompting is insufficient to override these entrenched patterns, underscoring the need for compositional analysis to diagnose latent risks in generative speech.

Figures & tables

Figure 1: Overall Framework. Two-stage evaluation demonstrated with PromptTTS++ model. Stage 1 establishes univariate gender priors, where an isolated descriptor like nurse triggers a strong female bias ( $P(\mathbf{x})=0.99$ ). In Stage 2 , recombining tokens with attributes like high-status and reckless creates a binding effect to the original female-leaning nurse , shifting the perceived gender toward male ( $P(\mathbf{x})=0.17$ ).

Model	Backbone	Text Encoder	Par. (Total/Text)
Vox (VoxInstruct)	LLaMA (AR + NAR)	mT5-base	$\sim$ 7B / 580M
Pr++ (PromptTTS++)	Diffusion + MDN	BERT	$\sim$ 150M / 110M
Par-M (Parler-Mini)	AudioLM	Flan-T5-large	880M / 780M
Par-L (Parler-Large)	AudioLM	Flan-T5-large	2.3B / 780M

Table 1: Evaluation of ITTS Models. Comparison of generative backbones, text encoders, and parameter scales. Par. denotes the number of parameters.

Panel A: Axis-Level Summary
Axis	Sub. / Trait	n	Vox	Pr++	Par-L	Par-M
Sta.	H. Status	1	$0.72$	$0.00$	$0.83$	$0.72$
Sta.	L. Status	1	$0.37$	$0.12$	$0.25$	$0.75$
Car.	F-ln.	10	$0.73\pm 0.06$	$1.00\pm 0.01$	$0.99\pm 0.01$	$0.96\pm 0.02$
Car.	Mix.	6	$0.71\pm 0.04$	$0.45\pm 0.20$	$0.98\pm 0.02$	$0.88\pm 0.06$
Car.	M-ln.	11	$0.70\pm 0.06$	$0.03\pm 0.06$	$0.77\pm 0.21$	$0.63\pm 0.15$

Table 2: Univariate Main Effect. Panel A summarizes axis-level marginal female probabilities (

P(\mathbf{x}_{\text{uni}})

) as Mean

\pm

SD. Panel B highlights exemplar traits. Abbreviations: Sta. (Status), H. (High), L. (Low), Car. (Career), F/M-ln. (Female/Male-leaning), Mix. (Mixed Prior), and Per. (Persona).

Sta. $\times$ Car.
Int. Cond.		Vox	Pr++	Par-L	Par-M
H. Sta.	+ F-ln. Car.	$0.86$	$0.40$	$0.91$	$0.95$
H. Sta.	+ M-ln. Car.	$0.78$	$0.20$	$0.74$	$0.80$
L. Sta.	+ F-ln. Car.	$0.73$	$0.69$	$0.62$	$0.89$
L. Sta.	+ M-ln. Car.	$0.68$	$0.39$	$0.29$	$0.54$
Sta. $\times$ Per.

Table 3: Bi-dimensional Interaction Analysis (

P(\mathbf{x}_{\text{bi}})

). See Table 2 for abbreviations.

Int. Cond.			Vox	Pr++	Par-L	Par-M
Sta.	Car.	Per.
High	F-ln. Car.	F-ln. Per.	$0.90$	$1.00$	$0.95$	$0.85$
	F-ln. Car.	M-ln. Per.	$0.85$	$0.17$	$0.91$	$0.88$
	M-ln. Car.	F-ln. Per.	$0.79$	$0.59$	$0.79$	$0.76$
	M-ln. Car.	M-ln. Per.	$0.80$	$0.00$	$0.79$	$0.62$
Low	F-ln. Car.	F-ln. Per.	$0.78$	$1.00$	$0.66$	$0.78$

Table 4: Tri-dimensional Interaction Dynamics (

P(\mathbf{x}_{\text{multi}})

). See Table 2 for abbreviations.

Bi-dimensional Interactions ( $\mathbf{x}_{\text{bi}}$ )
Int. Cond.	Vox	Pr++	Par-L	Par-M
Sta. $\times$ Car.
H. Sta. + F-ln. Car.	$-0.52$	$-0.41$	$-3.88$	$-2.60$
H. Sta. + M-ln. Car.	$-0.12$	$+7.81$	$+0.78$	$+1.35$
L. Sta. + F-ln. Car.	$+0.14$	$-1.81$	$-3.01$	$-3.61$
L. Sta. + M-ln. Car.	$+0.83$	$+6.14$	$+1.54$	$-0.04$

Table 5: Interaction Terms (

\mathcal{I}

). Log-odds deviation from additive baseline. Color intensity marks significance (Sec. 3.3 ). See Table 2 for abbreviations.

Axes ( $\mathcal{A}_{k}$ )	mT5-base	BERT-base	Flan-T5-large
$\mathcal{A}_{\text{sta}}$ (H / L)	-0.02 / -1.41	-5.06 / -0.60	0.77 / 0.47
$\mathcal{A}_{\text{car}}$ (F-ln. / Mix. / M-ln.)	0.58 / -0.90 / 0.36	1.67 / -0.48 / -1.71	1.65 / 0.45 / 4.16
$\mathcal{A}_{\text{per}}$ (O / C / E)	2.23 / 3.76 / 1.74	-1.15 / -1.60 / -0.50	1.17 / 0.36 / 1.49
(A / N)	2.06 / 2.98	-0.89 / -2.09	1.78 / 3.21

Table 6: Evaluated Text Encoders. Standardized bias scores and effect sizes (Cohen’s

d

Text-to-Speech

Text-to-speech converts written language into spoken audio, improving intelligibility, naturalness, voice control, multilingual coverage, and efficiency.

See all 67 papers

Training-Free Instruction TTS Gender Bias Calibration Using Model-Adaptive Steering
Oct 7, 2026Kuan-Yu Chen, Yi-Cheng Lin, Jeng-Lin Li +1Text-to-Speech TTS Synthesis
AI Research Center, Inventec Corporation, Taiwan · Graduate Institute of Communication Engineering, National Taiwan University, Taiwan
Voice or Stereotype? Disentangling Acoustic and Content-Based Gender in Speech-to-Speech Models
Sep 8, 2026Xiaoqun Liu, Tanu Mitra, Harshit Rajgarhia +1Speech Generation Evaluation Speech Processing
Centific Research · University of Washington
The Voice Behind the Words: Quantifying Intersectional Bias in SpeechLLMs
Mar 15, 2026Shree Harsha Bokkahalli Satish, Christoph Minixhofer, Maria Teleki +5Intersectional Bias in Language Models Language Model Bias Evaluation
Department of Speech, Music and Hearing, KTH Royal Institute of Technology, Sweden · Centre for Speech Technology Research, University of Edinburgh, UK · Texas A&M University, USA

Oct 7, 2026J/K move · Enter open · S save

Training-Free Instruction TTS Gender Bias Calibration Using Model-Adaptive Steering

Kuan-Yu Chen, Yi-Cheng Lin, Jeng-Lin Li, Jian-Jiun Ding

AI Research Center, Inventec Corporation, Taiwan · Graduate Institute of Communication Engineering, National Taiwan University, Taiwan

Text-to-Speech TTS Synthesis Representation Steering

Open paper AskPDF arXiv

Instruction text-to-speech (ITTS) systems systematically encode acoustic gender skews from descriptive style prompts, such as occupations or personas, even when demographic attributes are left unspecified. Calibrating the implicit gender distribution of the synthesized voices, across heterogeneous architectures and without model retraining or overriding explicit user prompts, is an open problem. In this work, we propose model-adaptive steering, a training-free bias calibration method that steers post-encoder conditioning representations using a group deviation vector paired with a coarse-to-fine strength search on a development set. A deterministic lexical gate bypasses intervention whenever explicit gender keywords are detected, preserving intended prompt semantics. Evaluated on a 12,800-prompt held-out benchmark across four ITTS models (three architectures), the proposed method reduces aggregate calibration error from 11.5--28.1 to 0.8--5.9 percentage points (e.g., shifting Parler-TTS Mini from 78.1% and PromptTTS++ from 27.3% female to 50.8--52.4%). Output rates stay within 0.9 pp across anchor set sizes at fixed operating points, while UTMOS decreases by at most 0.08 and WER degrades by at most 3.3 pp.

Abstract

Text-to-Speech

Explore similar work

Training-Free Instruction TTS Gender Bias Calibration Using Model-Adaptive Steering

Voice or Stereotype? Disentangling Acoustic and Content-Based Gender in Speech-to-Speech Models

The Voice Behind the Words: Quantifying Intersectional Bias in SpeechLLMs