Phase unwrapping based packet loss concealment using deep neural networks

Abstract

Recently, deep neural networks (DNN) has been employed to conduct a kind of packet loss concealment (PLC) method for digital speech transmission. Due to good mapping ability of the DNN, the DNN-based PLC method usually can achieve better speech recovery than some traditional methods. However, because the phase of speech is limited between - π and π and there is no obvious spectral structure like magnitude spectrum, it is not suitable for the DNN learning. The DNN-based PLC method is usually difficult to accurately estimate the phase of the lost speech, which also limits the quality of the recovered speech. In order to solve the problem of inaccurate phase estimation on the DNN, a new PLC method based on phase unwrapping is proposed in this paper. This method is divided into two stages: training stage and test stage. In the training stage, we first employ cellular automata (CA) to unwrap the phase of speech for making its spectrum structural. Then the unwrapped phase spectrum is used as the input feature and the training target to train the DNN model. The input feature of the DNN is consisted of the unwrapped phase spectra of few frames in front of the lost speech, and the training target is the unwrapped phase spectrum of the lost speech. In the test stage, firstly, the unwrapped phase of few frames in front of the lost speech is extracted as the input features of the DNN to obtain the unwrapped phase of the lost speech. Then the unwrapped phase of the lost speech is re-wrapped into the phase of the lost speech. Finally, combining with the estimated phase spectrum and logarithmic power spectrum (LPS) of the lost speech, we can recover the lost speech by the PLC. Experimental results show that, compared with the existing DNN-based PLC method, the proposed PLC method can better recover the lost speech and improve the quality of speech, which is suitable for the PLC of digital speech transmission.