基于SI-SDR优化的生成对抗网络语音增强方法

Speech enhancement method based on SI-SDR optimization generative adversarial network

为了解决现有的生成对抗网络(GAN)语音增强方法模型训练不稳定、生成语音质量不高的问题,提出一种尺度不变信号失真比(SI-SDR)优化的相对平均生成对抗网络(Ra GAN)语音增强方法.首先,构建一个基于生成对抗网络的端到端语音增强模型;然后,在模型中加入相对平均判别器,将真实数据和生成数据得分的差值作为模型训练的参考,显著增强了模型训练的稳定性;最后,采用SI-SDR直接度量生成语音的质量得分,并改进生成器训练的损失函数,将提高生成语音质量作为模型优化的目标.实验结果表明:相比基线方法,该方法可以有效提高未知噪声和低信噪比条件下的语音增强性能,增强后的语音具有更好的听觉质量和可懂性.

To solve the problems of unstable model training and low quality of generated speeches in the existing speech enhancement methods based on generative adversarial network(GAN),a speech enhancement method based on relativistic average GAN(RaGAN) optimized by scale-invariant signal-to-distortion ratio(SI-SDR) was proposed.First,an end-to-end speech enhancement model based on generative adversarial network was constructed.Then,a relativistic average discriminator was added to the model,and the difference value between the real data and generated data scores was used as a reference for the model training,which could make the model training more stable.Finally,the SI-SDR was used to measure the quality of the generated speeches and improve the loss function of generator training,which regarded improving speech quality as the goal of model training.Experiment results show that compared with the baseline methods,the proposed method can effectively improve the speech enhancement performance under noise unknown conditions and low signal to noise ratio conditions,and the enhanced speeches have better auditory quality and intelligibility.