Feedback utterances for computer-adied language learning using accent reduction and voice conversion method

This paper considers the generation of feedback utterances for speaking skills training of non-native English learners. The proposed feedback is in the form of a combination of the learner's voice and the linguistic gestures, i.e., the prosody or pronunciation, of a native speaker. Both accent reduction method and voice conversion method are employed to generate feedback stimuli. For accent reduction, three speech synthesis methods, namely pitch-synchronous overlap and add (PSOLA), harmonic stochastic model (HSM), and speech transformation and representation by adaptive interpolation of weighted spectrogram (STRAIGHT) are used to reduce the accent of the utterances of English learners. For voice conversion, the teacher's voice is converted to that of the learner and the converted speech is used as a feedback. Objective measurements are employed to assess the nativeness and acoustic quality of the generated stimuli. A feedback scheme which combines the accent reduction and voice conversion methods is also proposed.