结合高效通道注意力机制的语音增强算法仿真  

Simulation of Speech Enhancement Algorithm Combined with High-Efficiency Channel Attention Mechanism

作　　者：杨立东[1] 曾江蛟 董桂官[2] YANG Li-dong;ZENG Jiang-jiao;DONG Gui-guan(Inner Mongolia University of Science&Technology school of Information Engineering,Baotou Inner Mongolia 014010,China;China Electronics Standardization Institute,Beijing 100176,China)

机构地区：[1]内蒙古科技大学信息工程学院,内蒙古包头014010 [2]中国电子技术标准化研究院,北京100176

出　　处：《计算机仿真》2023年第3期258-262,535,共6页Computer Simulation

基　　金：国家自然科学基金项目(62161040);内蒙古自然科学基金项目(2021MS06030);内蒙古科技计划项目(2021GG0023)。

摘　　要：语音增强能有效解决人机交互和语音识别等技术中的噪声干扰问题。为了提高语音增强的效果,提出了一种结合高效通道注意力机制的生成对抗网络,并在两种不同的数据集下进行实验。该方法通过在生成对抗网络的生成器中加入高效的通道注意力机制来提高抑制无关于语音增强的信息来提高模型的灵活度与准确率。该模型在Nonspeech-100数据集相较于基线模型下语音感知质量评估(PESQ)平均提升了2.79%,语音短时客观可懂度(STOI)平均提升了0.95%;在NoiseX-92数据集下,ESQ平均提升了3.8%,STOI平均提升了2.03%。实验结果表明,该方法在没有增加很大的计算量的情况下提高了模型的性能。Speech enhancement can effectively solve the problem of noise interference in technologies such as human-computer interaction and speech recognition.In order to improve the effect of speech enhancement,a generative adversarial network combined with an efficient channel attention mechanism was proposed,and experiments were conducted on two different datasets.This method improves the flexibility and accuracy of the model by adding an efficient channel attention mechanism to the generator of the generation confrontation network to improve the suppression of information that is not related to speech enhancement.Compared with the baseline model under the Nonspeech-100 dataset,this model has an average improvement of 2.79%in speech perceptual quality assessment(PESQ),and an average improvement in speech short-term objective intelligibility(STOI)by 0.95%;under the NoiseX-92 dataset,PESQ increased by an average of 3.8%,and STOI increased by an average of 2.03%.Experimental results show that this method improves the performance of the model without increasing the amount of calculation.

关 键 词：语音增强 生成对抗网络 通道注意力 神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]