Researchers at Arizona State University found that Scarlett Johansson’s voice does indeed sound like “Sky,” OpenAI’s currently discontinued GPT-4o voice, according to NPR, which commissioned the comparison. After using AI models developed to analyze vocal similarities to compare Sky to roughly 600 other actresses, the lab reportedly found Johansson’s voice was “more similar to Sky than 98% of the other actresses.”
But the models also “often” found that the voices of Anne Hathaway and Keri Russell resembled Sky’s more than did Johansson’s. Visar Berisha, the professor who led the analysis, told NPR that Johansson’s voice is “similar, but likely not identical” to that of Sky.
One similarity was reportedly that if Sky had a vocal tract — the throat, mouth, and nasal passages that shape a given voice — it would be just as long as Johansson’s. But the models also found that Sky’s voice was a little higher-pitched and more expressive than Johansson’s, while Johansson’s is “slightly more breathy” than that of OpenAI’s model, the outlet writes.
Berisha told The Verge via email that the “black box” nature of AI models like those used in this analysis makes it “difficult to understand precisely what vocal similarities and differences the model is identifying” or why they might deem one voice more similar than another. Berisha’s other work includes OriginStory (PDF), an FTC challenge-winning microphone that watermarks voice recordings as human-created. OpenAI did not immediately respond to a request for comment.
Both OpenAI CEO Sam Altman and CTO Mira Murati have denied that Sky is intended to sound like Johansson. After its GPT-4o demo earlier this month — after which Altman published a one-word post that simply read “her” — Johansson said that Altman had asked her to lend her voice to the model, which she declined, and that he tried again as recently as two days prior to the demo.
Update June 2nd, 2024, 8:07AM ET: Added comment by ASU professor Visar Berisha and clarified wording on comparisons between Sky and Johansson.
