语音信号的分块稀疏表示分类研究

Investigation of Voice Signal Classification with Block Sparse

传统稀疏表示分类算法(SRC)在处理复杂多维的向量的时候,需要对稀疏后的每个信号单独处理求残差,会导致处理时间过长,无法有效地运用于实际的工程应用中。为解决这一问题,提出将图像处理的分块稀疏应用于语音稀疏表示分类的方法。该方法在传统稀疏表示分类的基础上,引入分块稀疏思想,将语音信号按指定的长度处理,从而将若干个稀疏系数组成稀疏组来进行进一步分类识别。验证实验表明,源于图像处理的分块稀疏表示分类法同样适用于语音信号的处理。实验结果表明,在识别率接近的情况下,语音信号分类识别所花费的时间比图像处理明显降低。这是因为图像稀疏分类的系数之间相关性较强,因而分类的识别率较高;而语音信号是典型的非平稳过程,各种特征参数随时间快速变化,因而根据长度分类的相关性显著减少。因此,语音信号识别的准确率虽然会有所降低,但其效率显著提升。

In dealing with complex multidimensional vector, traditional Sparse Representation Classification （SRC） has spent too much time on computing the residual error by sparse signal after each individual treatment, which is unable to be applied in practical engineering effectively. In order to solve this problem, the block sparse method of image processing has been introduced to the voice of the sparse rep- resentation classification, which is based on the traditional sparse representation classification and merged with idea of block sparse. Audio signal is treated via given length so that a sparse group has been constructed with several sparse coefficients for further classification in the voice field. Validation experiments have been conducted and its results show that block sparse representation classification stemmed from image processing can be applied in speech signal processing and that time consumption of audio signal classification is less than image processing under the condition of the same recognition rate, due to high correlativity among the coefficients of image sparse classification and thus high recognition rate. This is also because speech signal is classical non-stationary process and its characteristic parameters vary with time rapidly ,thus the correlativity of classification with length has been reduced significantly. Therefore,although accuracy of speech signal recognition could decrease, recognition efficiency would be exalted notably.