连续语音识别中声学建模的组合聚类算法研究

A Combined Clustering Algorithm of Acoustic Modelling for Continuous Speech Recognition

基于三音子连续语音识别的一个关键问题是在有限训练数据的条件下对大量声学模型参数的鲁棒性估计。为了解决这个问题 ,有两个主要的上下文相关的聚类算法被提出 ,它们是合并 (AgglomerativeClustering)聚类 (AGG)和决策树 (Tree based)聚类 (TB)。本文分析了这两种算法的优缺点 ,并分别对其进行了改进 ,然后提出了最大似然框架下组合聚类算法。大词汇量连续语音识别 (LVCSR)的实验结果表明 ,和单一的决策树聚类算法比较 ,提出的组合聚类算法对识别率有显著的提高。

A crucial issue in triphone-based continuous speech recognition is the large number of parameters to be estimated against the limited availability of training data. To cope with the problem, two major context-clustering methods, agglomerative (AGG) and tree-based (TB), have been widely investigated. We analyze both algorithms with respect to their advantage and disadvantage, develop several methods to improve on them, and introduce a novel combined method in the maximum likelihood framework. For LVCSR, the experimental results show the performance can be much improved by using the proposed combined method, compared with those of the existing TB method alone.