论文部分内容阅读
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 %~ 1 0 %.
To study the context-dependent vector quantization technique for continuous speech recognition, the k-means algorithm is proposed to adjust each context contained in the decision tree leaves one by one to optimize the mixture density of the phoneme model.Experimental results It is shown that the average density of context-dependent vector quantization using k-means algorithm is 4% -100% higher than the average distribution density obtained from simple merger decision tree leaves.