协调语音能量区域的正则化优化算法  

Regularization optimization algorithm for coordinating speech energy region

作　　者：师晨康 薛珮芸 白静 赵建星 SHI Chen-kang;XUE Pei-yun;+;BAI Jing;ZHAO Jian-xing(College of Information and Computer,Taiyuan University of Technology,Jinzhong 030600,China;Post-Doctoral Research Station,Shanxi Academy of Advanced Research and Innovation,Taiyuan 030032,China)

机构地区：[1]太原理工大学信息与计算机学院,山西晋中030600 [2]山西高等创新研究院博士后科研工作站,山西太原030032

出　　处：《计算机工程与设计》2024年第7期2173-2179,共7页Computer Engineering and Design

基　　金：山西省应用基础研究计划基金项目(201901D111094);山西省基础研究基金项目(青年)(20210302124544);山西省留学回国人员科技活动择优基金项目(20200017)。

摘　　要：为有效解决语音识别模型过拟合问题,提出一种协调语音能量区域的正则化优化算法。根据语音的共振峰特性,对语音信号高能量区域进行集体失活处理,增加模型对语音信号低能量区域的关注度;为进一步提升声学模型性能,采用堆叠8层的门控卷积神经网络提取语音时序特征,并对其中的门控机制进行优化,缓解梯度衰减现象;采用联结时序分类算法以汉字为建模单元对语音识别模型进行训练和解码。在公开中文语音数据集Aishell-1上的实验结果表明,该语音识别模型字错率降低至11.27%,与基线模型相比,字错率下降了7.93%,验证了该方法的有效性。To effectively solve the overfitting problem of the speech recognition model,a regularized optimization algorithm for coordinating speech energy regions was proposed.The high-energy areas of the speech signal were collectively dropped according to the resonance peak characteristics,increasing the model’s focus on the low-energy areas of the speech signal.To further improve the acoustic model performance,a gated convolutional neural network(GCNN)with stacked eight layers was used to extract speech timing features,and the gating mechanism in it was optimized to alleviate the gradient fading phenomenon effectively.The connectionist temporal classification(CTC)algorithm was used to train and decode the speech recognition model with Chinese characters as the modeling unit.Experimental results on Aishell-1,an open Chinese speech dataset,show that the word error rate of the speech recognition model is reduced to 11.27%,and the word error rate is reduced by 7.93%compared with the baseline model,which verifies the effectiveness of the method.

关 键 词：语音识别 声学模型 语音能量区域 正则化 卷积神经网络 联结时序分类 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]