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

2026-10-07

AI Summary

Instruction-based text-to-speech (ITTS) systems exhibit implicit gender biases that are difficult to calibrate without retraining. This work proposes a model-adaptive steering method that directly adjusts post-encoder representations for training-free bias correction through group bias vectors, coarse-to-fine intensity search, and deterministic lexical gating. To our knowledge, this mechanism is the first to operate without retraining while generalizing across heterogeneous architectures, eliminating implicit biases while strictly preserving explicit prompt semantics. Experimental results demonstrate that the proposed approach reduces aggregate calibration error by 0.8 to 5.9 percentage points across four mainstream ITTS models, with negligible degradation in audio quality and intelligibility.

Abstract

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 \textit{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.

Problem

• Instruction Text-to-Speech
• Gender Bias
• Bias Calibration
• Training-Free

Innovation

• Training-Free Bias Calibration
• Model-Adaptive Steering
• Instruction Text-to-Speech
• Gender Bias
• Deterministic Lexical Gate