Automatic Detection of Foreign Accent for Automatic Speech Recognition

Recognition of foreign accented speech remains among the most difficult tasks in automatic speech recognition. It was observed that using models trained on foreign data together with native models improves the recognition for speakers with foreign accent. However such an approach degrades the recognition performances on native speakers. In order to avoid such performance degradation the degree of accent should be detected prior to the recognition process. In this paper an automatic method of detection of the degree of foreign accent is proposed and results are compared with accent labeling carried out by an expert phonetician. This made possible a better targeting of speakers having a heavy foreign accent which allowed using the foreign accent dedicated model when necessary and thus improving recognition performances on non-native speech without major performance degradation on native speakers.