Abstract: After searching and analyzing the error forms of spoken initials, a two-time-two-level objective evaluation algorithm of initials was advanced on the basis of the phonetic knowledge. Ninety-eight combinations of initial and final were summarized as the basic elements of the initial objective evaluation. The experiment has manifested that the accuracy of the two-time-two-level algorithm is 2.56% higher than that of the Hidden Markov Model (HMM) algorithm, and 3.65% higher than that of BP neural network algorithm, and also 1.42% higher than that of single-time-two-level algorithm. The results prove that the calculation amount of decreases, and the accuracy increases.

Key words: speech objective evaluation, initial, speech signal processing, Mandarin proficiency test, computer assisted language learning

摘要： 通过罗列分析声母读音错误的主要表现形式，提出了利用语音知识引导的两层两级声母客观评测算法。根据普通话声母的特点，总结出98种普通话声韵组合作为声母的评测基元。实验结果表明，所提出的算法比单独用隐马尔可夫模型（HMM）算法评测的主客观符合率高2.56%，比单独用BP神经网络算法评测的主客观符合率高3.65%，比只用单层算法评测的主客观符合率高1.42%，证明该算法不但能减少计算量，还能提高评测的精度。

关键词: 语音客观评测, 声母, 语音信号处理, 普通话水平测试, 计算机辅助语言学习