摘要
针对汉语语音单音节结构的特点,考虑音节间协同发音的现象,本文提出了一种对三音子模型进行分级聚类的方法。与传统的基于决策树的状态聚类算法相比,该方法通过对稀少三音子模型聚类,更充分地利用训练数据,减少稀少三音子对状态聚类的影响,从而提高声学模型的鲁棒性。实验结果表明:大词汇量连续语音识别器采用这种分级聚类方法,不仅可以大大减少模型及其参数的数量,还可使系统识别率有所提高,其中误识率相对于传统的决策树状态聚类系统降低了4.93％。
Based on the single syllable characteristics of Mandarin and considering the inter-syllable coarticulatory phenomena, a new hierarchical clustering algorithm is proposed. Compared with the traditional decision-tree based state-tying, the algorithm can take better use of training data and lessen the impact of rare triphones to state-tying. Experiments on large vocabulary continuous Mandarin speech recognition system show that the method can get better performance (about 4.93% word error rate reduction) with even fewer parameters.
出处
《信号处理》
CSCD
2004年第5期497-500,共4页
Journal of Signal Processing
基金
上海市科委重点基金项目资助(01JC14033)
