摘要
针对DCNN网络缺乏直观的物理声学特征描述等问题,为了提高其在说话人识别系统中的识别性能,提出基于MFCC加权与改进深度卷积神经网络的说话人识别算法。算法首先提取信号的MFCC特征,并对特征进行分量凸显和加权改进,以提高特征中对最终识别准确率贡献大的分量的作用,然后通过改进深度卷积神经网络的结构并增加深度残差网络,进一步对准帧间信息,提高网络对说话人识别需求的适应性。实验结果表明,文中算法在不同的分段信噪比下均取得最优的识别准确率。
In response to the lack of intuitive physical acoustic feature descriptions in the DCNN network,to improve its recognition performance in the speaker recognition system,a speaker recognition algorithm based on MFCC weighting and improved deep convolutional neural network is proposed.The MFCC features of the signal is first extracted and component is highlighted,and weighting improvements is performed on the features to improve the role of the components in the features that have a large contribution to the final recognition accuracy.And then,by improving the structure of the deep convolutional neural network and adding a deep residual network,the inter-frame information is further aligned,and the adaptability of the network to the speaker recognition needs is improved.Experimental results show that the algorithm proposed in this paper achieves the best recognition accuracy under different segmented signal-to-noise ratios.
作者
倪美玉
曹为刚
NI Meiyu;CAO Weigang(Department of Electronic Information, Zhejang Vocational College of Science and Trade, Jinhua 321019, China)
出处
《微型电脑应用》
2022年第6期145-148,共4页
Microcomputer Applications
关键词
说话人识别
梅尔频率倒谱系数
特征加权优化
深度卷积神经网络
深度残差网络
speaker recognition
Mel frequency cestrum coefficient
feature weighting optimization
deep convolutional neural network
deep residual network