Application of Uni-Directional Microphone Array for Identifying English Pronunciation Errors

To identify the English pronunciation errors made by Chinese learners, this paper utilizes uni-directional microphones to construct a superdirective beamformer for capturing high quality input speech, and integrates the techniques of anti-model and confidence measure into the speech recognizer for accurate identification of the speaker's pronunciation errors. As to the beamformer, although designing a superdirective beamformer using omni-directional microphones is widely reported, little work details the designing using uni-directional microphones. We integrate the transfer function of the uni-directional microphones into the signal model of the beamformer, and derive the expression of the superdirective beamformer under the diffuse noise assumption. As to the speech recognizer, an anti-model for each phone is trained from the training data excluding the tokens of that phone. By integrating these anti-models into the recognition network, the recognizer can align the user's speech with the prompted text more accurately. Confidence measure is utilized to judge whether a segment of utterance is mis-pronounced. To justify the proposed techniques, simulated noises are injected into utterances provided by some Chinese undergraduates in Nankai University. Recognition results are compared with the judgment made by several English linguistic experts. Experiment result shows the effectiveness of the proposed beamforming algorithm and the recognition techniques.