Speech enhancement using microphone array neural switched Griffiths-Jim beamformer

There is a great need for speech enhancement in today´s world due to the increasing demand for speech based applications. These applications vary from hearing-aids, hands-free telephony to speech controlled devices. The main goal is to minimize the interference from an acquired speech signal. The interference we considered here could be from any noise source such as competing speaker, radio, TV and so on. This paper proposes a solution to improve the current design of the switched Griffiths-Jim beamformer structure. It introduces an adaptive nonlinear neural network algorithm 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. Comparison analysis of the traditional four channel linear beamformer and the proposed four-channel 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 surrounding. The nonlinear approach introduced here shows remarkable improvement over the previous linear adaptive beamformer approach.