Underwater acoustic source separation with deep Bi-LSTM networks

The underwater acoustic source separation task has great practical significance, while it remains a challenge due to the high cost of collection and annotation of the datasets. Currently, underwater acoustic source separation is mainly based on the theory of Blind source separation. In this paper, we proposed a time-frequency domain source separation method, in which the deep bidirectional Long Short-Term Memory (Bi-LSTM) recurrent neural networks (RNNs) are utilized to process the short-time Fourier transformation (STFT) magnitude features and to estimate the ideal amplitude mask (IAM) target. Furthermore, the mean square error (MSE) between the labeled masks and the estimated masks are calculated as the cost function. Our underwater acoustic source separation model is discussed and evaluated on the dataset of ShipsEar. The signal-level evaluation indicator and the application-level indicator are employed to assess our method. Experimental results suggest that the proposed method is a meaningful exploration and attempt for this task.