Text-Independent Foreign Accent Classification using Statistical Methods

In this paper we present an automated approach for text-independent foreign accent classification. Results of foreign accent classification task could be used for adapting acoustic models, modifying lexicon, changing language model with regards to non-native speakers. In our study, we investigate statistical approaches which differ from the a priori knowledge they need: GMM, which requires neither phonetic knowledge nor labelling, phone recognition (without a lexicon), sentence recognition (with a lexicon and a grammar). This work is done in the framework of the HIWIRE (Human Input that Works In Real Environment) European project. We evaluated the proposed approaches on English speech corpus pronounced by French, Italian and Greek speakers. All experiments were performed in speaker-independent and text-independent mode. The best classification rate (83.3%) is achieved using sentence recognition or a phone-based recognition followed by language modelling approach.