Original article was published on Artificial Intelligence on Medium
Challenges we faced
Duplicating the unique accent of Bryant is a huge challenge. Unlike most of the celebrities, the pronunciation and accent of Kobe Bryant are different from the others. The phoneme varies between different dialects of English. However, it is quite difficult to find suitable resources that can accurately translate Bryant’s accent into the correct phoneme. Luckily, though there are limited resources, we finally managed to find a suitable one and solve the problem.
Lack of suitable voice data is another huge challenge for us. Normally speaking, we could simply record one’s voice to extract its audio dataset. However, as we are cloning Bryant’s voice this time, searching suitable dataset becomes an important task. Unlike the other actors, having few clips with clear audio, we have to pay a relatively long time to do the research work. Since most of the audio is extracted from interviews, the casual environment causes a higher level of contamination, for example, cough and laughter. With limited choices, we have to edit the clips to maintain the clearness of the dataset.