Context-Dependent Vector Quantization Based on K-Means Algorithm
-
摘要: 研究用于连续语音识别的语境相关矢量量化技术.提出采用k-means(k-均值)算法 逐一地调整决策树叶子所包含的各个语境,实现对音素模型的混合密度的优化.实验结果表 明,采用k-means算法的语境相关矢量量化得到的平均分布密度比简单合并决策树叶子所得 到的平均分布密度提高4％～10％.Abstract: An approach based on k-means algorithm to implement context-dependent vector quantization for continuous speech recognition is presented. With the approach, individual phonetic context of each leaf in a decision tree can be moved from one leaf to another ,in order to optimize the mixture density given by that tree. Experimental results given in this paper demonstrate that the average likelihood given by this approach based on k-means algorithm is four to ten percent higher than that obtained with the previous method of optimizing mixture densities by merging leaves of decision trees.
-
计量
- 文章访问数: 2733
- HTML全文浏览量: 74
- PDF下载量: 1164
- 被引次数: 0
下载:
