基于波束形成的长短时记忆网络语音分离算法研究

在利用深度学习方式进行语音分离的领域，常用卷积神经网络(RNN)循环神经网络进行语音分离，但是该网络模型在分离过程中存在梯度下降问题，分离结果不理想。针对该问题，该文利用长短时记忆网络(LSTM)进行信号分离探索，弥补了RNN网络的不足。多路人声信号分离较为复杂，现阶段所使用的分离方式多是基于频谱映射方式，没有有效利用语音信号空间信息。针对此问题，该文结合波束形成算法和LSTM网络提出了一种波束形成LSTM算法，在TIMIT语音库中随机选取3个说话人的声音文件，利用超指向波束形成算法得到3个不同方向上的波束，提取每一波束中频谱幅度特征，并构建神经网络预测掩蔽值，得到待分离语音信号频谱并重构时域信号，进而实现语音分离。该算法充分利用了语音信号空间特征和信号频域特征。通过实验验证了不同方向语音分离效果，在60°方向该算法与IBM-LSTM网络相比，客观语音质量评估(PESQ)提高了0.59，短时客观可懂(STOI)指标提高了0.06，信噪比(SNR)提高了1.13 dB，另外两个方向上，实验结果同样证明了该算法较IBM-LSTM算法和RNN算法具有更好的分离性能。

Research on Long Short-Term Memory Networks Speech Separation Algorithm Based on Beamforming

In the field of speech separation using deep learning, the Recurrent Neural Network (RNN) is commonly used for speech separation, but the network model has a gradient descent problem in the separation process, and the separation result is not ideal. Considering this problem, this paper uses Long Short-Term Memory (LSTM) network to explore the signal separation, which makes up for the deficiency of RNN network. The separation of multi-channel vocal signals is more complicated, and most of the separation methods used at this stage are based on the spectrum mapping method, and the spatial information of the voice signal is not effectively used. In response to this problem, this paper combines the beamforming algorithm and the LSTM network to propose a beamforming LSTM algorithm. The voice files of three speakers are randomly selected from the TIMIT voice library, and the super-pointing beamforming algorithm is used to obtain beams in three different directions. The spectral amplitude characteristics in each beam are extracted, and a neural network is constructed to predict the masking value. The to-be-separated speech signal spectrum is obtained. and the time-domain signal is constructed, and the speech separation is realized. The algorithm makes full use of the spatial characteristics of the speech signal and the signal frequency domain characteristics. The effect of speech separation in different directions is verified through experiments. Compared with the IBM-LSTM network, at 60-degree direction, this algorithm improves Perceptual Evaluation of Speech Quality (PESQ) by 0.59, Short-Time Objective Intelligibility (STOI) index by 0.06, and Signal to Noise Ratio (SNR) by 1.13 dB. At the other two reverse directions, the experimental results also prove that the algorithm has better separation performance than the IBM-LSTM algorithm and the RNN algorithm.