The zero-crossing rate is an important feature in the recognition of speech. However, the feature is not very reliable in the computation. A method of conversion of the zero-crosssing rate into the main period and rec...The zero-crossing rate is an important feature in the recognition of speech. However, the feature is not very reliable in the computation. A method of conversion of the zero-crosssing rate into the main period and recognition of Putonghua initials by HMM based on the main period are proposed in this paper. The study of on-line recognition of aspirate voiceless stop consonants P, T and K, unaspirate affricative consonants ZH, Z and fricative consonant H of Chinese based on the new method has been made and we obtained 78 percent of the correct recognition rate for the initials. The implementation technique of the new method and the experiment of the initial recognition are described in the paper.展开更多
为了进行连续马尔可夫模型的初值提取,提出一种各类在训练样本空间近似均衡分布的K均值聚类法。在聚类的过程中引入惩罚因子,从而限制过多的训练矢量集中于一个或几个类,使样本空间划分近似均匀。连续马尔可夫模型初值提取实验证明,该...为了进行连续马尔可夫模型的初值提取,提出一种各类在训练样本空间近似均衡分布的K均值聚类法。在聚类的过程中引入惩罚因子,从而限制过多的训练矢量集中于一个或几个类,使样本空间划分近似均匀。连续马尔可夫模型初值提取实验证明,该方法与标准的K均值聚类法、LBG(L inde Buzo G ray)聚类法相比,降低了矢量量化产生的全局失真,各个类在样本空间的分布更加均匀,提高了矢量量化的性能。将该方法用于孤立词识别连续马尔可夫模型的初值提取,可使各个高斯概率密度函数的参数估计更逼近其无偏估计,从而提高了马尔可夫模型初值的可靠性。展开更多
文章分析了经典隐马尔可夫模型（Hidden Markov Model，HMM）齐次假设的理论缺陷，以及两种非齐次HMM。语音识别对比实验表明，经验性的惩罚概率法是稳健的、且更有效的补偿方法。实验结果还指出在最优惩罚概率下，经典HMM达到了与非齐...文章分析了经典隐马尔可夫模型（Hidden Markov Model，HMM）齐次假设的理论缺陷，以及两种非齐次HMM。语音识别对比实验表明，经验性的惩罚概率法是稳健的、且更有效的补偿方法。实验结果还指出在最优惩罚概率下，经典HMM达到了与非齐次的基于段长分布的HMM（Duration Distribution Based HMM，DDBHMM）几乎相同的识别率，证明了齐次假设并不影响经典HMM在实用中的重要性。文章提出了一种改进Baum－Welch重估算法的初值的经验方法，用于HMM参数的估计，在汉语连续语音识别实验中一致性地降低了音节误识率。展开更多
文摘The zero-crossing rate is an important feature in the recognition of speech. However, the feature is not very reliable in the computation. A method of conversion of the zero-crosssing rate into the main period and recognition of Putonghua initials by HMM based on the main period are proposed in this paper. The study of on-line recognition of aspirate voiceless stop consonants P, T and K, unaspirate affricative consonants ZH, Z and fricative consonant H of Chinese based on the new method has been made and we obtained 78 percent of the correct recognition rate for the initials. The implementation technique of the new method and the experiment of the initial recognition are described in the paper.
文摘为了进行连续马尔可夫模型的初值提取,提出一种各类在训练样本空间近似均衡分布的K均值聚类法。在聚类的过程中引入惩罚因子,从而限制过多的训练矢量集中于一个或几个类,使样本空间划分近似均匀。连续马尔可夫模型初值提取实验证明,该方法与标准的K均值聚类法、LBG(L inde Buzo G ray)聚类法相比,降低了矢量量化产生的全局失真,各个类在样本空间的分布更加均匀,提高了矢量量化的性能。将该方法用于孤立词识别连续马尔可夫模型的初值提取,可使各个高斯概率密度函数的参数估计更逼近其无偏估计,从而提高了马尔可夫模型初值的可靠性。
文摘文章分析了经典隐马尔可夫模型（Hidden Markov Model，HMM）齐次假设的理论缺陷，以及两种非齐次HMM。语音识别对比实验表明，经验性的惩罚概率法是稳健的、且更有效的补偿方法。实验结果还指出在最优惩罚概率下，经典HMM达到了与非齐次的基于段长分布的HMM（Duration Distribution Based HMM，DDBHMM）几乎相同的识别率，证明了齐次假设并不影响经典HMM在实用中的重要性。文章提出了一种改进Baum－Welch重估算法的初值的经验方法，用于HMM参数的估计，在汉语连续语音识别实验中一致性地降低了音节误识率。
