一种改进的DNN瓶颈特征提取方法

Modified DNN Bottleneck Feature Extraction Method

针对梅尔频谱倒谱、梅尔标度滤波器组特征提取、瓶颈特征等常用语音特征对语音前后帧相关性信息提取不足、冗余信息较多导致识别率不高的问题,对此改进了深度神经网络的语音瓶颈特征提取方法。该方法利用重叠组套索算法增强表征力、L2,1范数稀疏正则化对深度神经网络语音瓶颈特征提取进行改进,经过梅尔频谱倒谱声学特征提取进行二次特征提取,获得表征能力较强、具有稀疏性的语音瓶颈特征。实验结果表明,改进的深度神经网络语音瓶颈特征与原始的梅尔频谱倒谱特征提取方法相比,语音识别错误率降低了3.25%。

The common speech features such as Meyer spectrum cepstrum,Meyer scale filter set feature extraction,and bottleneck feature are insufficient to extract the frame correlation information before and after the speech,and more redundant information leads to low recognition rate.Aiming at these problems,the speech bottleneck feature extraction method of deep neural network is modified.This modified method uses the overlapping group lasso algorithm to enhance the representation,and the L2,1 norm sparse regularization to improve the speech bottleneck feature extraction of deep neural networks.After the second feature extraction by Meyer spectral cepstral acoustic feature extraction,the speech bottleneck feature with strong representation ability and sparseness is acquired.The experimental results indicate that the speech recognition error rate of the improved deep neural network is reduced by 3.25% as compared with the original Mel spectrum cepstrum feature extraction method.