Recognition of Non-Native Speakers

We propose a novel and promising non-native speaker recognition system by using both high-level language cues and low-level acoustic accent cues. The system consists of a high-level language cue based sub-system and four accent recognition based sub-systems. Each sub-system can be used individually or together via system fusion. The novelty and uniqueness of our proposed approach are as follows: First, it utilizes both high-level language cues and low-level accent cues, which will largely improve the robustness; Second, it employs a new concept of accent signature, that is to be learned in a completely data-driven fashion, which guarantees portability across different accents; Third, it applies various system fusion approaches which guarantees improved overall performance.