基于Conformer-SE的端到端语音识别  

End-to-end Speech Recognition Based on Conformer-SE

作　　者：马永杰 李罡 MA Yong-Jie;LI Gang(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin 132022,China;School of Mechanical and Control Engineering,Baicheng Normal University,Baicheng 137000,China)

机构地区：[1]吉林化工学院信息与控制工程学院,吉林132022 [2]白城师范学院机械与控制工程学院,白城137000

出　　处：《计算机系统应用》2024年第12期106-114,共9页Computer Systems & Applications

基　　金：2022年度吉林省教育厅科学技术研究项目(JJKH20220013KJ);2023年大学生创新创业训练计划(202310206035)。

摘　　要：基于自注意力机制的Transformer端到端模型在语音识别任务中表现出了卓越的性能.然而,该模型在浅层处理时对局部特征信息的捕捉能力存在一定的局限,同时也没有充分考虑不同块之间的相互依赖性.为了解决这些问题,提出了一种改进的Conformer-SE端到端语音识别系统模型.该模型首先采用了Conformer结构来替代Transformer中的编码器部分,从而增强了模型对局部特征的提取能力.接着,通过引入SE注意力通道机制,将每个块的输出以加权求和的形式整合到最终的输出中.在Aishell-1这一公开数据集上的实验结果显示,相较于原始的Transformer模型,Conformer-SE模型在字符错误率上相对降低了18.18%.The end-to-end Transformer model based on the self-attention mechanism shows superior performance in speech recognition.However,this model has limitations in capturing local feature information during shallow processing and does not fully consider the interdependence between different blocks.To address these issues,this study proposes Conformer-SE,an improved end-to-end model for speech recognition.The model first adopts the Conformer structure to replace the encoder in the Transformer model,thus enhancing its ability to extract local features.Next,by introducing the SE channel attention mechanism,it integrates the output of each block into the final output through a weighted sum.The experimental results on the Aishell-1 dataset show that the Conformer-SE model reduces the character error rate by 18.18%compared to the original Transformer model.

关 键 词：语音识别 端到端 TRANSFORMER CONFORMER SE注意力通道 

分 类 号：TN912.34[电子电信—通信与信息系统]