基于残差-通道注意力卷积神经网络的高效语音降噪算法研究
摘要
融合残差连接与通道注意力机制的卷积神经网络(RCACNN)的语音降噪算法。该算法通过残差结构缓解梯度
消失问题,利用通道注意力增强关键特征提取。基于TIMIT和NOISEX-92数据集的实验表明,在0dB信噪比下,
RCACNN较最优基线的信噪比增益达6.5dB(提升2.4dB),同时短时客观可懂度达到0.86(提升0.11),总体参数量
减少18%,单条处理仅需8.8ms。研究成果为智能语音交互提供了高效解决方案。
关键词
全文:
PDF参考
[1]Schröter H, Escalante-B A N, Rosenkranz
T, et al. DeepFilterNet: A low complexity speech
enhancement framework for full-band audio based on
deep filtering[C]//2022 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP). IEEE,
2022: 7407-7411.
[2]Lu Y X, Ai Y, Ling Z H. MP-SENet: A speech
enhancement model with parallel denoising of magnitude and
phase spectra[C]//Proc Interspeech 2023. 2023: 3834-3838.
37
计算机与通信进展 | 第2卷/第10期
Advances in Computers and Communications
DOI: 10.21437/Interspeech.2023-1441.
[3]Zheng L, Pan H, Kong L. Ripple attention for visual
perception with sub-quadratic complexity[C]//International
Conference on Machine Learning. PMLR, 2022: 26993-27010.
[4]Lan C, Zhang L, Zhang Y, et al. Attention
mechanism combined with residual recurrent neural network
for sound event detection and localization[J]. Journal of
Audio, Speech, and Music Processing, 2022, 29: 1-12. DOI:
10.1186/s13636-022-00263-6.
[5]徐浩森,姜囡,齐志坤。基于自注意力机制的
卷积循环网络语音降噪[J].计算机仿真,2024,41(4):
500-506. DOI:10.3969/j.issn.1006-9348.2024.04.095.
[6]Zhang L, Wang M, Zhang Q, et al. Environmental
attention-guided branchy neural network for speech
enhancement[J]. Applied Sciences, 2020, 10(3): 1167. DOI:
10.3390/app10031167.
[7]Zhou W, Robertson G, Sjöstrand H.深度异质联合
结构:燃料性能代码的时间频率替代模型[J].核能年鉴,
2025,211:110893.
[8]郑丹,宋思楠,贺强。深度学习中深度可分离卷
积网络在遥感图像分类中的优化[J].信息系统工程,2024
(10):132-135. DOI:10.3969/j.issn.1001-2362.2024.10.035.
Refbacks
- 当前没有refback。