基于改进MP稀疏分解的语音识别方法

在限定输入词汇量的条件下进行语音识别的过程中,结合MP稀疏分解的原子结构特性,提出把语音信号经MP稀疏分解所得的最佳原子时频参数作为匹配参数对语音进行识别。用基于遗传算法和原子库划分的策略对MP稀疏分解的寻优过程进行改进以提高MP稀疏分解的效率。在Matlab环境下进行仿真试验时,采用LGB算法对测试信号进行判别。试验结果表明,MP稀疏分解算法经改进后运行速度得到提高,采用10个原子的时频参数可有效识别长度约为6000的语音信号。

Speech recognition based on improved MP sparse decomposition algorithm

Based on the limit of input glossaries and the characters of atoms structure in Matching Pursuit (MP) sparse decomposition, a new speech recognition algorithm was proposed through using the best obtained atoms' time-frequency parameters as matching parameters among test signals and swatch signals. In order to enhance the computing efficiency of MP sparse decomposition, the strategy based on dictionary partitioning and Genetic Algorithm (GA) was applied. In Matlab, the LGB algorithm was used to distinguish test signals. The simulation results show that the speech signal with length 6000 can efficiently be recognized through using ten atoms' time-frequency parameters, while the running speed of improved MP sparse decomposition algorithm is enhanced.