Voice cloning study improves speaker similarity without worsening accent

Researchers report that a new method for cross-language voice cloning preserved more of a speaker's identity at a comparable accent level in tests across several open speech models. Their training-free technique estimates an accent pattern from a synthetic voice rendered in two languages, then subtracts that pattern while generating speech. On OmniVoice, the reported speaker-similarity gain above the usual identity-accent trade-off was +0.11 to +0.27 across three test sets. A twelve-listener panel and a separate language-identification measure supported the finding. The method showed no gain on one tested model, X-Voice.