Microphone Array Speech Separation Algorithm based on DNN

Microphone array speech separation in high reverberant and low signal-to-noise ratio (SNR) environments, is still a challenge. This paper utilizes a deep neural network (DNN) and spatial information of a microphone array to train ideal binary masks and ideal ratio masks of target signals, then utilizes these masks to separate the target speech signals in a noisy and reverberant environment. In sound localization, SRP-PHAT is robust to noise and reverberation, while the SRP-PHAT computation is time-consuming in real application. To alleviate this problem, GSRP-PHAT which combines SRP-PHAT with Gammatone filter banks is proposed. GSRP-PHAT implying spatial information is applied as the input feature for the DNN. The simulation results show that the proposed algorithm achieves better performance in terms of source-to-distortion ratio, source-to-interference ratio, and short-time objective intelligibility in low signal to noise ratio (SNR) and high reverberant environments for omnidirectional speech separation. Also, the proposed method can still maintain the performance under the untrained conditions.