Separation of Multiple Stationary Sound Sources using convolutional neural network
Swapnil G. Mali, Mohit V. Dhale, Shriniwas P. Mahajan · 2021
This paper presents a deep neural network-based approach for separating multiple sound sources located at different positions in-room environments. The Mixed signal is recorded using a Linear microphone array in-room environment. We have also shown that the known signals are correctly separated with the Convolutional Neural Network-based algorithm using mapping mixed-signals with the original signal. We have proposed a novel separation architecture and database creation algorithm with various combinations of gaussian noise source signals to create a diverse mixture for training and testing purposes. After the sound separation model's training and testing, it is shown that the system's performance is better compared with other sound separation models. It is beneficial to separate the multiple mixed signals in a known source case compared to an unknown source case.