Sound Source Separation Using Spatio-temporal Sound Pressure Distribution Images and Machine Learning

Sound source separation (SSS) using a microphone array is effective in various situations, such as in the recording of target speech in noisy environments. We previously proposed an SSS system using a differential-type array and time-delay neural network (NN). An advantage of the differential-type array was that distortion of the target speech due to the nonlinear property of the NN was prevented. However, the system was only effective for a narrowband signal. For broadband signals (such as speech), a new SSS system is proposed, which extends the previous one in the following two areas. First, the input to the NN is extended to sound pressure distribution images, which are formed based on the microphone outputs. Second, the number of layers in the NN is increased. Computer simulations revealed that the proposed system exhibited higher SSS performance than conventional arrays, including our previous one.