Multi-microphone speech enhancement technique using a novel neural network beamformer : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Engineering at Massey University, Albany, New Zealand
This thesis presents a novel speech enhancement algorithm to reduce the background noise from the acquired speech signal. It introduces an innovative idea for the speech beamformer using an input delay neural network based adaptive filter for noise reduction.
Speech communication is considered as the most popular and natural way for humans to communicate with computers. In the past few decades, there has been an increased demand for speech-based applications; examples include personal dictation devices, hands-free telephony, voice recognition for robotics, speech-controlled equipment, automated phone systems, etc. However, these applications require a high signal-to-noise ratio to function effectively. The background noise sources such as factory machine noises, television, radio, computer or another competing speaker, often degrade the performance of the acquired signals. The problem of removing these unwanted signals from the acquired speech signal has been investigated by various authors. However, there is still room for improvement to the existing methods.
A multi-microphone neural network based switched Griffiths-Jim beamformer structure was implemented using the Labview software. The conventional noise reduction section of the Griffiths and Jim beamformer structure was improved with a non-linear neural network approach. A partially connected three-layer neural network structure was implemented for rapid real-time processing. The error back-propagation algorithm was used here to train the neural network structure. Although it is a slow gradient learning algorithm, it can be easily replaced with other algorithms such as the fast back-propagation algorithm.
The proposed algorithms show promising noise reduction improvement over the previous adaptive algorithms like the normalised least mean squares adaptive filter. However, the performance of the neural network depends on its chosen parameters such as learning rate, amount of training given, and the size of the neural network structure. Tests with a speech-controlled system demonstrate that the neural network based beamformer significantly improves the recognition rate of the system.