基于CHF-CNN的语音分离

Speech Separation Based on CHF-CNN

深度神经网络已经在语音分离方面取得很好的表现,但是卷积神经网络获取的语音信息会更全面。经常用来评估预测目标好坏的分类准确率和命中率-错误率(HIT-FA)之间存在不平衡现象。为了解决这种不平衡,对卷积神经网络的损失函数进行了改进,提出使用二元交叉熵及命中率-错误率混合(CHF)损失函数,构成CHF-CNN模型。实验证明,使用CHF-CNN模型可以同时提高分类准确率和命中率-错误率(HIT-FA)来避免不平衡现象。此外,还验证了不同信噪比下的语音分离成果,发现当信噪比匹配时效果比不匹配时明显好,同时随着信噪比的增大效果会越来越好。

Deep neural networks have achieved good performance in speech separation,but the speech information obtained by convolutional neural networks will be more comprehensive.There exists an imbalance between the classification accuracy rate and HIT-FA which are often used to assess the quality of predicted targets.In order to solve this imbalance,we improved the loss function of the convolutional neural network and proposed to use the binary cross-entropy HIT-FA hybrid(CHF) loss function to form the CHF-CNN model.Experiments have shown that using the CHF-CNN model can avoid imbalance by simultaneously improving classification accuracy and HIT-FA.In addition,we also verified the results of speech separation under different SNR.It is found that when the signal-to-noise ratio is matched,the effect is better than that of the mismatch,and the effect will be better as the signal-to-noise ratio increases.