There are two immediate flaws I see with the study. The first is the sound quality of the youtube videos, which have been engineered to human hearing. Dogs hear frequencies high above the human range, and it's highly likely their vocalizations carry nuance and context in the upper frequency. A simple microphone recording won't do - you need something with high sensitivity and minimal distortion, and you have to ensure the digitization of the sound file doesn't interfere with the quality for those higher ranges.
The second is the sound input into the analysis, which also should be tuned for these higher frequencies - I find it unlikely that the AI was matching frequencies above human hearing to generate dog "words".
In essence, they're analyzing how humans hear dogs speak, not how dogs speak to each other. To do so properly, an ideal research setting would be one of dogs interacting live with each other, with sound recording equipment that can match the quality of canine hearing.
That being said, it's a fascinating work, and I understand that the researchers were working with the data on hand - definitely shows promise for continued study.