Accent Analogy Guidance: More Speaker Similarity at Equal Accent in Cross-Lingual Voice Cloning

Accent Analogy Guidance raises cross-lingual TTS speaker similarity at equal accent across several open models.

Accent Analogy Guidance (AAG) is a training-free sampler term for cross-lingual zero-shot TTS that subtracts an accent direction estimated from one synthetic voice rendered in both languages, so the voice cancels and only the accent remains. Authors score methods by speaker similarity above the identity-accent trade-off of classifier-free guidance reweighting at equal accent (ΔSIM). Across open models, OmniVoice gains ΔSIM +0.11 to +0.27 (accent 3.51–4.28 on a 1–5 scale at similarity 0.29, versus 0.02 for reweighting); MaskGCT, CosyVoice 2, and F5-TTS also lie above their curves, while a premise test correctly predicts no gain on X-Voice. A blind LLM accent judge, an LLM-free language-ID measure, and a twelve-listener panel agree.