Traditional acoustic parameters detect commercial but not open-source voice clones: a paired study of Italian speakers

Preprint, not peer reviewed. Voice cloning has already been used in fraud and challenges the evidential value of recorded speech. This study tests whether routinely measured forensic phonetics parameters can distinguish a real voice from its clones. Ten native Italian speakers (five women, five men) read ten sentences with forensically plausible content twice, and a short story once; each was cloned with a commercial service (ElevenLabs, v4) and an open-source model run locally (XTTS-v2). Vowel duration, F0, F1, F2 and F3 were measured on 870 hand-corrected stressed vowels per channel, while F0 variability, jitter, shimmer, HNR and CPPS were measured on 300 sentences, both as clean signals and after Opus compression at 16 kbit/s. Both clones produced shorter vowels and an expanded vowel space (median vowel triangle area +61% and +39%). The commercial clone had a higher F0 in all ten speakers (+18.7 Hz), more variable intonation and greater regularity, including a higher CPPS; the open-source clone stayed close to the speakers’ F0 and was slightly less regular. Prosodic differences clearly exceeded the variation between two readings by the same speaker, and the effects survived compression. A logistic-regression classifier using six acoustic-phonetic metrics normalised against each speaker’s reference readings achieved an AUC of 0.91 for the commercial clone but only 0.63–0.70 for the open-source clone. Traditional analysis, therefore, remains informative, but the open-source clone challenges its reliability: it largely evades these measures by staying close to the speaker’s pitch and regularity, and is exposed mainly by expert listening.