Corpus-based modeling of naturalness estimation in timing control for non-native speech

In this paper, aiming at automatic estimation of naturalness in timing control of non-native's speech, we have analyzed the timing characteristics of non-native's speech to correlate with the corresponding subjective naturalness evaluation scores given by native speakers. Through statistical analyses using English speech data spoken by Japanese with temporal naturalness scores ranging one to five given by natives, we found high correlation between their scores and the differences from native's speech. These analyses provided a linear regression model where naturalness in timing control is estimated by differences from native's speech in durations of overall sentences, individual content and function words and pauses. The proposed naturalness evaluation model was tested its estimation accuracy using open data. The root mean square errors 0.64 between scores predicted by the model and those given by the natives turned out to be comparable to the differences 0.85 of scores among native listeners. Good correlation between model prediction and native's judgments confirmed the appropriateness of the proposed model.