
Research Article
A Complex Neural Network Adaptive Beamforming for Multi-channel Speech Enhancement in Time Domain
@INPROCEEDINGS{10.1007/978-3-030-99200-2_11, author={Tao Jiang and Hongqing Liu and Yi Zhou and Lu Gan}, title={A Complex Neural Network Adaptive Beamforming for Multi-channel Speech Enhancement in Time Domain}, proceedings={Communications and Networking. 16th EAI International Conference, ChinaCom 2021, Virtual Event, November 21-22, 2021, Proceedings}, proceedings_a={CHINACOM}, year={2022}, month={4}, keywords={End-to-end Multi-channel Speech enhancement Complex operations}, doi={10.1007/978-3-030-99200-2_11} }
- Tao Jiang
Hongqing Liu
Yi Zhou
Lu Gan
Year: 2022
A Complex Neural Network Adaptive Beamforming for Multi-channel Speech Enhancement in Time Domain
CHINACOM
Springer
DOI: 10.1007/978-3-030-99200-2_11
Abstract
This paper presents a novel end-to-end multi-channel speech enhancement using complex time-domain operations. To that end, in time-domain, Hilbert transform is utilized to construct a complex time-domain analytic signal as the training inputs of the neural network. The proposed network system is composed of complex adaptive complex neural network beamforming and complex fully convolutional network (CNAB-CFCN). The real and imaginary parts (RI) of the clean speech analytic signal are used as training targets of the CNAB-CFCN network, and the weights of the CNAB-CFCN network are updated by calculating the scale invariant signal-to-distortion ratio (SI-SDR) loss function of the enhanced RI and clean RI. It is fundamentally different from the complex frequency domain single channel approach. The experimental results show that the proposed method demonstrates a significant improvement in end-to-end multi-channel speech enhancement scenarios.