Dupree is a Washington-based political journalist and radio host for local broadcaster WSB Atlanta. He began to lose his voice in 2016 and was diagnosed with tongue protrusion dystonia, a neurological condition which causes people to lose control over their tongues, making speech almost impossible.
While Dupree continued to work as a journalist, losing his voice meant he had to come off the air. After a two-year absence, he will be back broadcasting this month with a new AI-generated voice on WSB Atlanta and other Cox Media-owned stations in Orlando, Jacksonville, Dayton, and Tulsa.
"It is me, there is no doubt about that," Dupree told the BBC. "Yes, it is slightly robotic, but no-one was promising me that it was going to be perfect."
Cereproc was founded in 2006 and specialises in text to speech technology. It has created over 250 voices for people who have lost the ability to speak. Often they lose their voice due to degenerative diseases like Motor Neurone Disease (ALS) or throat cancer.
"We take a big database of speech and cut it up and stitch it back together," cofounder and CTO of Cereproc Christopher Cox told Business Insider.
In the case of Dupree, Cereproc collected data from his previous shows and reports. This was a painstaking process, as Cereproc has to accurately transcribe any "ums" and "ers," as well as tiny mispronunciations.
Cox explained: "We take that text and that audio, we break it down into the sounds of the language. We don't use words as units, it's broken down into the vowels and consonant sounds that you have in English."
By training a neural network on these elements of speech, Cereproc then learns how Dupree produced those sounds, and was able to build a new voice.
It took Cereproc about two months to re-create Dupree's voice, but Cox said this was longer than usual because of the time it took to transcribe his speech. Most people come to Cereproc before they lose their voice and are able to record a script, which is converted into a voice within days.
Once the person's voice is built, they can install it on a computer and type whatever they wish to say, then it will play through speakers. The software can also be used with systems that track things like eye and muscle movement.
The advantage of such AI-generated voices is that people can have a voice that really sounds like them, and this even goes for people who have never spoken.
Cox told Business Insider about a young man who had never been able to speak. "He was from Newcastle and he really wanted a Geordie voice," Cox said. "He was just stuck with using quite posh received pronunciation in a British accent, or American accents, when he wanted to talk So one of his friends donated his voice."
Cereproc also used its technology to re-build the voice of President John F. Kennedy, and read the speech he never delivered in Dallas on the day he was assassinated.
Also from Business Insider South Africa: