Title: Examining the Acoustic and Perceptual Divergence Between Standard and Accented Speech and Their Synthesized Clones

While voice cloning technology is frequently assessed based on general audio quality, there is limited understanding regarding how well it maintains specific accents and the resulting impact on listener perception. This study employs a hybrid approach, integrating computational metrics with perceptual testing, to contrast standard Mandarin speech against heavily accented Mandarin, alongside their respective cloned versions.

Computational analysis utilizing embeddings revealed that the distance between original and cloned speech was greater for accented speakers across several speaker-discriminative embedding spaces. However, this disparity vanished when the data was normalized against the baseline variability observed within each speaker’s original recordings.

In the accompanying perceptual experiment, listeners rated the clones as more similar to their source voices when the source was a standard speaker compared to an accented one. Furthermore, intelligibility improved from the original speech to the clone in all cases, with the most significant gains observed in the accented speech samples.

These findings indicate that accent variations influence both the perceived identity match and the clarity of voice clones, even when such variations do not manifest in baseline-normalized speaker-embedding distances. Consequently, the results suggest that accent preservation should be regarded as a distinct element of speaker identity preservation, challenging the assumption that standard, off-the-shelf speaker-discriminative embeddings fully account for it.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC