基于瞬时相位差和深度学习的丢包隐藏方法

实时IP 语音通信在数据包会丢失的情况下,语音质量会受到严重影响。为了恢复传输过程中丢失的语音信息,本文提出了一种基于瞬时相位差(Instantaneous Phase Deviation, IPD)和深度神经网络(Deep Neural Network, DNN)的丢包隐藏 (Packet Loss Concealment, PLC)方法。在训练阶段,将语音的对数功率谱(Log Power Spectrum, LPS)和IPD作为训练DNN的输入特征,以学习从接收包到丢失包的映射关系;在重构阶段,将丢包前接收到的语音包送入训练好的DNN中,恢复出丢失包的语音。实验结果表明,在不同丢包率下,所提方法的性能优于传统的基于LPS和DNN的PLC方法。 

Packet loss concealment based on instantaneous phase deviation and deep neural network

In the case of packet loss for real-time speech communication, the speech quality will be seriously affected. In order to recover the lost speech information during transmission, this paper proposes a packet loss concealment (Packet Loss Concealment, PLC) based on instantaneous phase deviation (Instantaneous Phase Deviation, IPD) and deep neural network (Deep Neural Network, DNN). In the training stage, the log power spectrum (Log Power Spectrum, LPS) and IPD of the speech are used as the input feature of the DNN training for learning the mapping relationship from the received packets to the lost packets. In the reconstruction stage, the received packets are sent to the well trained DNN for recovering the lost packet. Experimental results prove that under different packet loss rates, the proposed algorithm can gain better performance than conventional LPS+DNN-based PLC method.