Article ID: e26.122
Voice cloning can retain speaker similarity without guaranteeing preservation of specific acoustic characteristics. We characterize ten Japanese speakers using natural Japanese, matched same-language synthetic Japanese, and cross-lingual synthetic English generated with CosyVoice 3. Speaker-related information was evaluated with system-aligned CAMPPlus and independent WavLM-SV embeddings, while F0 and HNR provided interpretable acoustic measures. Same-language synthesis retained strong speaker-related similarity but consistently reduced within-utterance F0 variability, while HNR changes were speaker-dependent and reproducible. Cross-lingual speaker-retention results were representation-dependent; a supplementary matched Japanese–English control using a bilingual speaker further showed feature-dependent acoustic effects across languages.