Mel频率下语音信号深度频谱特征提取方法仿真

Simulation of Depth Spectrum Feature Extraction Method for Speech Signal under Mel Frequency

采用当前方法提取语音信号频谱特征时,不能有效去除语音信号中存在的噪声信号,提取得到的特征误差较大,存在抗干扰性能差和特征提取结果准确率低的问题。针对上述问题,提出Mel频率下语音信号深度频谱特征提取方法。对噪声信号进行经验模态分解,将分解得到的IMF分量进行门限域处理,通过对应的滤波方案去除语音信号中存在的噪声信号。采用Mel滤波器处理去噪后的语音信号,得到Mel频率的语音信号。利用线性预测系数描述Mel频率下的语音信号,并对其进行微分处理,将微分向量进行加权处理,根据一定的加权比例重组微分向量,利用语音信号深度频谱特征参数,完成Mel频率下语音信号的深度频谱特征提取。仿真结果表明,所提方法的抗干扰性能高、特征提取结果准确率高。

The current method cannot effectively remove the noise signal in voice signal. The extracted feature error is large. The anti-interference performance is poor and the accuracy of feature extraction result is low. Therefore, a method to extract the depth spectrum feature of voice signal at Mel frequency was proposed. The empirical mode decomposition was performed on the noise signal. The decomposed IMF component was subjected to threshold domain processing. The noise signal existing in voice signal was removed by the corresponding filtering scheme. Mel filter was used to process the denoised voice signal, so as to obtain the Mel frequency. The linear prediction coefficient was used to describe the voice signal at Mel frequency and the voice signal was differentiated. The differential vector was weighted. According to a certain weighting ratio, the differential vector was reconstructed. Finally, the depth spectrum feature parameter of voice signal was used to complete the depth spectrum feature extraction of voice signal at Mel frequency. Simulation results show that the proposed method has high anti-interference performance and high accuracy of feature extraction result.