AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 09:18:24 UTC.

Briefings

A new training-free method called Accent Analogy Guidance (AAG) improves cross-lingual voice cloning by canceling the reference voice's accent while preserving identity. Experiments on four open-source TTS models show AAG consistently places speaker similarity above the identity-accent trade-off curve, with MaskGCT and CosyVoice 2 performing particularly well on F5-TTS. Results were validated by blind LLM judges and a twelve-listener panel, confirming the method's effectiveness across different model architectures.

Key points

• Accent Analogy Guidance (AAG) is a training-free technique that cancels the reference voice's accent while preserving speaker identity in cross-lingual text-to-speech generation.
• AAG consistently outperforms standard reweighting approaches, placing speaker similarity above the identity-accent trade-off curve on four open TTS models.

Context

This research addresses a critical issue in cross-lingual voice cloning where the reference voice's accent leaks into the target speech. The study proposes Accent Analogy Guidance (AAG), a training-free sampler term that subtracts an accent direction estimated from the model's own predictions for one synthetic voice rendered in both languages. This approach allows for precise accent cancellation while maintaining speaker identity, validated by blind LLM judges and a twelve-listener panel.

Limitations

Results are limited to four open TTS models (OmniVoice, MaskGCT, CosyVoice 2, and F5-TTS) and rely on a single blind LLM judge and a twelve-listener panel. The study does not include a comprehensive evaluation of all possible reweighting settings, and the findings may not generalize to closed-source models or different evaluation metrics.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 26/10/2026, 18:40:49 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibsbh6fxl4nmtrkov5f3x54clws4ca46vxg5ptko6rhhwpnptygci