基于深度自编码网络语音识别噪声鲁棒性研究

摘要

To solve the problem of the center and the radius determined by randomly in the speech recognition tasks based on traditional Radial Basis Function(RBF)neural network, an unsupervised pre-training method which uses a large number of unlabeled data to initialize the network parameters is proposed to replace the traditional random initialization method based on the layered mechanism of human brain on speech recognition. This paper introduces the Deep Auto-Encoder(DAE) neural network as acoustical model and further analyzes robustness of speaker-independent isolated speech recognition on small size vocabulary database. The experimental results show that DAE outperforms RBF with MFCC(Mel Frequency Cepstrum Coefficient)feature extraction. In addition, compared to MFCC, GFCC(Gammatone Frequency Cepstrum Coefficient)gives more attribution on anti-noise property with a relative accuracy improvement of 1.87%in collaborate with DAE network.%为了解决传统径向基(Radial basis function,RBF)神经网络在语音识别任务中基函数中心值和半径随机初始化的问题,从人脑对语音感知的分层处理机理出发,提出利用大量无标签数据初始化网络参数的无监督预训练方式代替传统随机初始化方法,使用深度自编码网络作为语音识别的声学模型,分析梅尔频率倒谱系数(Mel Frequen-cy Cepstrum Coefficient,MFCC)和基于Gammatone听觉滤波器频率倒谱系数(Gammatone Frequency Cepstrum Coefficient,GFCC)下非特定人小词汇量孤立词的抗噪性能.实验结果表明,深度自编码网络在MFCC特征下较径向基神经网络表现出更优越的抗噪性能;而与经典的MFCC特征相比,GFCC特征在深度自编码网络下平均识别率相对提升1.87%.