I would be very interested to see if providing speaker locale as a feature to a machine learning algorithm would increase its performance for speech recognition.