Tone Evaluation Algorithm Of Mandarin Continuous Speech

Tone is an important factor in mandarin speech, and the degree of accuracy of the tone is one important factor to determine good and bad in Putonghua. The tone evaluation system is an important part in Computer Assisted Language Learning (CALL) and PUTONGHUA SHUIPING CESHI (PSC). As to the continuous speech influenced by context, the continuity between the tones can not be ignored, which has a serious impact on the performance of the tone evaluation system. In order to raise the accuracy of the evaluation algorithm, this paper improve the characteristics of the traditional tri-syllable tone contour feature GMM model with considering the suprasegmental information in the continuous FO curve. The research results are as follows.1. Tracking FO curve:In continuous speech, a syllable’s tone is related to the adjacent syllables. But the tones between the two syllables are often ignored. To get the transition information from the previous voiced region to the current one or from the current to the next voiced region, the pitch value of unvoiced region is interpolated with Spline. Based on tri-syllable Spline interpolation, the correlation of the machine score and human score can achieve0.7309.2. The choice of tone characteristic:FO contour is considered to be the superposition of three components as the Fujisaki model of tonal language. These components are:the Phrase Component, which is affected by the intonation of the sentence. The Accent Component, which corresponds to lexical tone of the syllable. The base frequency, which is a constant for each speaker. Based on the Fujisaki-model, the paper extracts only the lexical tone from the FO contour to train GMM model. And the result shows that the correlation in test set is improved14.09%than the traditional feature.3. Improvement of tone model:To test and verify the transition information of the pitch is great helpful for the tone evaluation, this paper test the nucleus model and the transition information model. The result of the comparison is that the correlation of transition information model is better than the nucleus model. And compare to the0.7361from the Fujisaki model in2, the correlation based on nucleus is only0.6137. This result shows that the syllable frequency transition raise the accuracy of the tone evaluation.