Acoustic and perceptual differences between standard and accented speech and their voice clones

Researchers analyzed how voice cloning technology preserves accented speech compared to standard speech, finding that clones of accented speakers show larger perceptual differences from originals despite similar baseline-normalized embedding distances. The study reveals that accent variation significantly impacts perceived speaker identity and intelligibility in voice cloning systems, suggesting current speaker-discriminative embeddings don't fully capture accent preservation.

This research addresses a critical gap in voice cloning evaluation by examining accent preservation as a distinct component of speaker identity rather than an incidental feature. The study's dual computational and perceptual approach reveals a nuanced finding: while embedding-space measurements suggested larger original-clone distances for accented speech, this difference disappeared when normalized against speaker-specific variability. However, human listeners perceived clones as less similar to accented originals than to standard originals, indicating that subjective identity preservation diverges from computational metrics. The intelligibility gains from original to clone were larger for accented speech, suggesting the cloning process may inadvertently homogenize accent features while improving clarity. This tension between computational measures and perceptual outcomes highlights limitations in relying solely on off-the-shelf speaker embeddings for evaluating voice cloning quality. The findings carry implications for voice synthesis applications where accent fidelity matters—including language learning tools, accessibility applications, and cultural preservation contexts. Developers currently optimizing for speaker similarity using standard embeddings may unknowingly be degrading accent preservation, particularly for non-standard speech varieties. The research suggests the field needs explicit accent-aware metrics and training objectives rather than assuming general speaker embeddings capture this dimension adequately. Future work should investigate whether accent-specific optimization improves perceived identity preservation across diverse speech varieties.

• →Voice cloning systems preserve accented speech less effectively than standard speech in human perception, despite similar computational embedding distances.
• →Current speaker-discriminative embeddings fail to fully capture accent variation as part of speaker identity.
• →Cloning increases intelligibility more for accented than standard speech, suggesting potential accent homogenization during synthesis.
• →Accent preservation should be treated as an explicit optimization target rather than an assumed byproduct of speaker similarity.
• →Developers relying solely on standard embeddings for quality evaluation may inadvertently degrade accent fidelity in voice cloning applications.