Masked-speech Recognition Using Human and Synthetic Cloned Speech.
Lauren Calandruccio, Mohsen Hariri, Emily Buss and 1 others
PMID 41358594WHAT IT FOUND
Overall, synthetic voice clones were understood about as well as the human voices they copied, and listeners often could not tell them apart.
Some individual voices differed, and nonsense sentences were harder than meaningful ones.
Key findings
01Overall, human talkers and their synthetic clones were understood almost equally well, with 47% correct for human speech and 46% correct for clones.
02Meaningful sentences were understood better than nonsense sentences, with 57% correct versus 36%, but the human/clone difference changed with sentence type and voice.
03Listeners often judged a human voice and its clone to be the same person, and rated clones only slightly less human-like than the humans.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The study used five female talkers only, and the paper says results should be examined for male speech. The listeners were healthy young adults with normal hearing and native or comfortable English, so it does not show how people with hearing loss, children, older adults, or speakers of other dialects or accents would perform. The overall human/clone similarity hid voice-specific differences, with some human/clone pairs differing by more than 10 percentage points on recognition. Only 13% of sentences were repeated back correctly in their entirety at -6 dB SNR, so many responses were incomplete. The scoring comparison was made with clear, normal-hearing young adult speech, so it does not show how human or machine scoring would work for clinical speech errors, different accents, or child speech.
Declared interests
The work was supported by the College of Arts and Sciences and the Office of Faculty Advancement at Case Western Reserve University. The authors declared no potential conflicts of interest.
The easy way to misread this
Do not conclude that synthetic voice clones are interchangeable with human talkers for clinical speech testing. The overall similarity hid voice-specific differences, and the study used young adults with normal hearing producing clear English, so it does not show the same result for children, hearing loss, disordered speech, or other accents.