Voice or Stereotype? Disentangling Acoustic and Content-Based Gender in Speech-to-Speech Models
Xiaoqun Liu, Tanu Mitra, Harshit Rajgarhia, Abhishek Mukherji
cs.SD · cs.AI
Quick verdict
Claims S2S bias hides in gender attribution, not voice drift; 5 models misgender up to 90%; convincing, but one voice per gender and no code/data release temper it
Summary
The paper argues that standard fairness audits of speech-to-speech (S2S) models are structurally blind: because deployed S2S systems answer in one fixed output voice, probes that watch the output voice for stereotype drift can never fail. The authors run a 3×3×2 factorial study — English, Spanish, Mandarin; masculine/neutral/feminine-stereotyped LLM-written first-person passages; male/female Azure TTS voices — yielding 180 voiced passages screened by a carrier gate (0/60 classifier flips per language). Five S2S models (GPT-4o-audio, Gemini 2.5 Live, GLM-4-Voice, Kimi-Audio, Step-Audio-2-mini) complete five tasks scored on two channels: voice rendering and gender attribution. RQ1 finds no reliable voice drift for any model (pooled contrast −0.018±0.020), partly because fixed timbres make drift architecturally impossible. RQ2 finds every model's stated gender follows content: pooled odds ratios of 1.7–3.4 for open models versus 21.95 (GPT-4o-audio) and 24.27 (Gemini); misaligned cells misgender 83–100% of speakers (90% pooled) versus 2% in aligned cells. Summarize pronoun slopes track content (Δpp≈0.43–0.60 for the steepest models). Limitations: ten passages per cell, one voice per gender, single-rater validation, binary coding only, LLM-authored/LLM-rated passages, Gemini's Mandarin transcript produced by its own transcription service, and no released code or data.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Standard S2S fairness probes that monitor output-voice stereotype drift are structurally unable to detect bias in fixed-voice architectures, so a clean drift result is uninformative about fairness (§1, Contribution 1).
- All five tested S2S models shift their stated speaker-gender judgment with content stereotype while the true voice is held fixed: pooled odds ratios of 1.7–24.3 per content step, with the closed models largest (21.95 and 24.27) (Table 7).
- When voice and content clash, models misgender 83–100% of speakers depending on language (90% pooled), versus 2% when they agree, showing the acoustic evidence is available but overridden (Table 2, §4).
- The rendered output voice shows no reliable stereotype drift for any model (pooled Δ = −0.018 ± 0.020), and the content×voice interaction is nonsignificant in all fifteen readback fits (§4, Table 2).
- Language typology modulates the surface but never the direction: Mandarin, where spoken tā is gender-homophonous, is the only language in which no model stays voice-faithful (misaligned misgendering ≥20% for all five) (§4).
Strengths
Structural validity diagnosis of fixed-voice audits
Section 1 and Contribution 1 argue that a drift probe cannot fail on a single-timbre S2S system — its null is guaranteed by construction. This reframes prior speech-fairness evaluations (Spoken StereoSet, VoiceBBQ) as potentially uninformative rather than merely unrepresentative, echoing Lum et al. 2025's 'trick tests' critique.
Instrument validation before testing (carrier gate)
Section 3.4 and App. G audit all 180 inputs before any model hears them: hard criterion 0/60 classifier flips per language on two wav2vec2 classifiers; soft criterion F-tests on the femininity composite (p=.97/.18/.41 for en/es/zh). The lone sub-.05 channel (es, clf-1 margin, p=.018) is decomposed three ways (directional contrast −0.0009 logits, p=.54), a model of honest confound handling.
Clean factorial contrast with matched null and alternative channels
Large, coherently reported attribution effects
Separation-honest reporting
Weaknesses
One voice per gender per language limits the causal claim
Described audit protocol is not released
Attribution sensitivity vs. skill confound in the open models
LLM-generated, LLM-rated stimuli with a single human check
No formal multiple-comparison control
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This preprint (arXiv 2609.09263, Sep 2026; Centific Research + Univ. of Washington) asks whether speech-to-speech (S2S) models let the gender stereotype of spoken *content* override the *acoustic* gender of the speaker's voice. It splits the question into RQ1 (voice rendering: does the output voice drift toward the stereotype?) and RQ2 (gender attribution: when the model states the speaker's gender, does it follow voice or content?). The central contribution is a validity argument: because deployed S2S systems speak in one fixed output voice, drift-watching probes are structurally guaranteed to return nulls (Contribution 1, §1). The headline empirical claim is that all five tested models decide gender from content — pooled odds ratios of 1.7–24.3 per content-femininity step (Table 7), with misaligned cells misgendering 83–100% of speakers (90% pooled) versus 2% in aligned cells (Table 2).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of this paper about speech-to-speech models and gender?
The paper finds that speech-to-speech models judge a speaker's gender mainly from what is said (stereotyped content) rather than from how the voice sounds. When a speaker's voice and the stereotype of their words disagree, models misgendered 83–100% of such speakers across three languages (about 90% pooled), versus only about 2% when voice and content agree. The models' rendered output voice itself shows no stereotype-driven drift.
Why can't standard bias audits detect this problem?
Because most deployed speech-to-speech systems answer in one fixed, hard-coded output voice that never changes with the input. A bias test that watches the output voice for drift toward a stereotype can therefore never fail, even on a biased model — the paper calls this structural blindness and argues audits must instead test the model's spoken or written gender judgments, scored on invariance to topic changes.
Which models and languages were tested?
Five systems: OpenAI GPT-4o-audio, Google Gemini 2.5 Live, and three open-source models (GLM-4-Voice 9B, Kimi-Audio 7B, Step-Audio-2-mini 8B), across English, Spanish, and Mandarin. Inputs were 180 passages read by male and female Azure TTS voices. Closed models showed the largest content bias (odds ratios ≈22–24 per content step), though the most voice-faithful single judgment came from an open model (Kimi-Audio in English).
What should practitioners do differently after this paper?
The authors advise that a stable output voice is not evidence of gender fairness: before shipping captions, pronouns, or personas, test the system's attribution behavior with matched and clashing voice/content inputs scored on invariance, or use voice-preserving models. Mitigations to try include suppressing unsolicited gender inference, weighting the voice over content priors, and allowing refusal or uncertainty.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- CJK characters did not survive PDF text extraction (e.g., Mandarin prompts in App. D.1 and the tā homophone discussion appear with gaps); this does not affect any numeric claims reviewed here.
- The extraction ends naturally at Appendix H's closing text; no mid-document truncation was detected (references and appendices A–H are present in full).
- Any supplementary artifacts (passage lists, exact classifier outputs) may exist outside the PDF; their absence is asserted only with respect to the provided text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.