Improvement of Spatial Ambiguity in Multi-Channel Speech Separation Using Channel Attention

Multi-channel speech separation has been successfully applied in a complex real-world environment such as the far-field condition. The common solution to deal with the far-field condition is using a multi-channel signal captured by a structured microphone array and leveraging the inner difference between channels to enhance the speech separation performance. The spatial feature has been widely used in recent speech separation research. This feature appears to be insufficient when the location information becomes ambiguous. This is known as the spatial ambiguity problem. In order to deal with the spatial ambiguity problem, this study proposes an attention mechanism for the Temporal-Spatial Neural Filter (TSNF), in which the channel attention on merged features and the feature map of 1D convolution block in the temporal convolution network is proposed. The proposed method is evaluated on the multi-channel reverberant dataset which is built based on the WSJ0-2mix dataset. The dataset is simulated in the real-environment room by using the Room Impulse Response generator. In the experimental results, the proposed methods produced the SI -SNR improvement of about 1.2dB in close speakers' case, while a small decrease of 0.1dB in other cases.