Multi-channel speech coding has become an important research field due to the demand of high quality speech in real life. This paper proposes a novel multi-channel speech coding method based on deep neural networks (DNN) by employing sound localization called time difference of arrival (TDOA). As Figure 1 shown, the DNN and TDOA estimation are used to compress and recover multi-channel speech signals.

Nowadays, microphone array plays an important role in speech signal processing. This paper proposes a new parametric coding scheme to further improve the coding efficiency of linear microphone array. In the coding scheme, Inter-channel Level Difference (ICLD), Inter-channel Time Difference (ICTD) and Inter-channel Coherence (ICC) are used as spatial parameters. As shown in Figure 1, at the encoder, a down-mixed signal is obtained by averaging all signals time-aligned with the reference channel. Only the spatial parameters between the first channel and the last channel are extracted. At the decoder, the speech signal of each channel is reconstructed by the down-mixed signal received, spatial parameters and their interpolation.