Abstract: Speech signal dimensions are higher when the signal is transformed to frequency domain, manifold learning algorithm can find a smooth low-dimensional manifold embedded in the high-dimensional data space. The manifold learning algorithm is proposed to reduce the dimensions in the high-dimensional data for Mandarin digit speech recognition. Low-dimensional manifold structure is extracted from the high-dimensional frequency data based on locally linear embedding of manifold learning algorithms. Then the resulting low-dimensional data is inputted into Dynamic Time Warping（DTW） to recognize. Simulation results demonstrate that the dimensions are lower using Local Linear Embedding（LLE） compared with MFCC, the recognition rate increases by 1.2% in Mandarin digit speech recognition, and the recognition speed gets improved effectively.

Key words: Mandarin digit recognition, manifold learning algorithm, Locally Linear Embedding（LLE）, Dynamic Time Warping（DTW）

摘要： 语音信号转换到频域后维数较高，流行学习方法可以自主发现高维数据中潜在低维结构的规律性，提出采用流形学习的方法对高维数据降维来进行汉语数字语音识别。采用流形学习中的局部线性嵌入算法提取语音频域上高维数据的低维流形结构特征，再将低维数据输入动态时间规整识别器进行识别。仿真实验结果表明，采用局部线性嵌入算法的汉语数字语音识别相较于常用声学特征MFCC维数要少，识别率提高了1.2%，有效提高了识别速度。

关键词: 汉语数字识别, 流形学习, 局部线性嵌入算法, 动态时间规整