I just assumed that such people might be more interested in using the app and hence have more influence on defining/training the Turkish accent in the algorithm. In the end, it doesn’t really have to be a bias of course

I mean, if there’s a bias towards those who have been living abroad long enough to speak almost accent-free, the algorithm might detect minor details that we are not aware of.

Any chance the training data has been mostly provided by Turkish expats? Selection bias might be involved