Within-Speaker Features for Native Language Recognition in the Interspeech 2016 Computational Paralinguistics Challenge

The Interspeech 2016 Native Language recognition challenge was to identify the first language of 867 speakers from their spoken English. Effectively this was an L2 accent recognition task where the L1 was one of eleven languages. The lack of transcripts of the spontaneous speech recordings meant that the currently best performing accent recognition approach (ACCDIST) developed by the author could not be applied. Instead, the objectives of this study were to explore whether within-speaker features found to be effective in ACCDIST would also have value within a contemporary GMM-based accent recognition approach. We show that while Gaussian mean supervectors provide the best performance on this task, small gains may be had by fusing the mean supervector system with a system based on within-speaker Gaussian mixture distances