Investigation of Suitable Training Conditions for Nonlinear Beamformer

Optimization of Acoustic Beamformer in the Low Frequency for Noise Reduction

“…Neural Fig. 1: Neural Network Beamformer [10] networks are widely used for pattern recognition, data classification, etc. This neural network is also applied to nonlinear beamforming [10].…”

“…The RNN has an internal, directed closed path in the hidden layer, as shown in Fig. 2 [10]. This structure allows RNNs to store information and dynamically change their behavior temporarily.…”

“…The more complex the problem is, the longer the required length becomes. Theoretically, the entire past input history should be taken into account, but it is said that the output can only reflect at most 10 times in the past, so while short-term memory can be achieved, long-term memory is difficult to achieve [10].…”

“…The band-pass filtering of 300 Hz to 3400 Hz simulates a traditional telephone band. The band-limited signals are convolved with impulse responses measured with a microphone array [10].…”

“…In the previous study, a neural network-based beamformer was trained and optimized by using wideband speech signals. The alternative NN-based beamformer was trained with mixtures of sinusoidal signals [10]. However, the size of the microphone array to which the sound sources are input has not been successful in learning in the low-frequency range.…”