基于特征比较和模拟退火-遗传算法的普通话音素评分模型

为了帮助发音困难者障碍者和外语学习者矫正普通话发音错误,提出基于Mel频率倒谱系数(Mel frequencycepstrum coefficient,MFCC)特征比较和模拟退火-遗传算法(simulated annealing genetic algorithm,SAGA)的普通话音素评分模型。该模型采用动态时间弯折(dynamic timewarping,DTW)算法对普通话音素进行相似度比对,并基于SAGA评分机制对发音进行自动评分。本文对比了不同优化算法(SAGA和局部优化算法)、不同DTW算法对语音评分的影响。结果发现:SAGA评分模型下的音素评分正确率大于94%,远远优于局部优化算法。此外,在SAGA评分模型下,搜索路径为平行四边形的改进DTW算法具有最优的评分结果。因此,基于MFCC和SAGA的评分模型适用于普通话音素评分。

Pronunciation scoring model for Mandarin Phonemes based on feature comparison using a simulated annealing genetic algorithm

A Mandarin Phoneme pronunciation scoring model was developed to help people with difficulty in pronunciation and people learning foreign languages to rectify pronunciation errors.The method uses feature comparison of the Mel frequency cepstrum coefficient(MFCC) and a simulated annealing genetic algorithm(SAGA).The dynamic time warping(DTW) algorithm is used to evaluate the phoneme similarity,and to automatically compute the scores for these phonemes based on the SAGA scoring mechanism.This paper compares phoneme scores using different optimization algorithms(SAGA and local optimization) and different DTW algorithms.The results show that the SAGA model accuracy is better than 94%,significantly better than the local-optimization model.Moreover,the combination of SAGA and the improved DTW algorithm with a parallelogram search path resulted in the best pronunciation score.Thus,the model based on the MFCC and SAGA methods is appropriate for pronunciation scoring of Mandarin Phonemes.