Automatic Tone Assessment for Strongly Accented Mandarin Speech

Abstract

This paper discusses the tone scoring part of a Mandarin pronunciation scoring system. It recognizes tones of isolated syllables and words by using a GMM model and uses the recognition results for tone assessment. Initially, experiment results are bad on strongly accented speech. There are two reasons: one is that the inaccurate force-alignment leads to incomplete FO contours; the other is due to the special patterns of FO contours. We propose three solutions. The first is to make the extraction of FO contour independent of the force-alignment. The second is to base the scoring on GMM posterior probabilities. The third is to use the same accented speech to train the GMM model. After these improvements are taken, the tone scoring correct rate is improved form 60.2% to 83.1% and the final average score difference between machine and human's evaluations is decreased from 16.77 to 6.43