Multi-microphone adaptive neural switched Griffiths-Jim beamformer for noise reduction
Vaitheki Yoganathan, Tom James Moir · 2010
This paper proposes a novel multi-microphone nonlinear neural network based switched Griffiths-Jim beamformer structure for speech enhancement. The main objective of this algorithm is to reduce real-world interference signals such as radio, television or computer fan noise from an acquired speech signal. The proposed algorithm improves the current design of the switched Griffiths-Jim beamformer structure by introducing an adaptive nonlinear neural network filter for the noise reduction section. The network topology used here is a partially connected three-layer feedforward neural network structure. The error backpropagation algorithm is used here as the learning algorithm. A comparison analysis of the traditional three-microphone linear beamformer and the proposed three-microphone neural switched Griffiths-Jim beamformer structure is discussed here. They are both tested with different types of interference signal from the Noise-X database. All the experiments are conducted in real-world surroundings. The nonlinear approach introduced here shows remarkable improvement over the previous linear adaptive beamformer approach.