Comparing with Mel-frequency cepstral coefficients (MFCC), the sound feature based on Power-Law Nonlinearity and Power-Bias Subtraction ( PNCG) has good ability in overcoming the noise. First, MFCC is combined with PNCC to form hybrid eigenmatrix, and the contrast experiments on hybrid feature and traditional feature are conducted with HMM, SVM and GMM model respectively. Secondly, the HMM which has obtained better results is chosen to filter the test samples in advance, and the SVM and GMM are then selected for the secondary classification, and the recognition accuracy of two kind of two-layer models are tested as well. Experimental results show that the use of HMM/GMM two-layer model and mixture feature can achieve preferable recognition effect in noisy environment.%相比Mel倒谱系数(MFCC),基于能量偏差移除和幂函数的声音特征(PNCC)具有较强的抗噪能力.首先,将PNCC和MFCC组成混合特征矩阵,在隐马尔科夫模型(HMM)、高斯混合模型(GMM)和支持向量机(SVM)下对混合特征和传统特征做对比实验.其次,先选取实验结果较好的HMM模型过滤测试样本,再分别选取GMM和SVM做二次分类,并测试两种双层模型的识别正确率.结果表明在噪声环境下使用HMM/GMM双层模型和混合特征可取得较好的识别效果.
展开▼
