An analysis of adaptive neural networks for speech enhancement

Adaptive algorithms have the versatile characteristics when applied on signals. In this paper, we attempt to enhance the speech signal using adaptive techniques. Initially, adaptive linear neuron (ADALINE) model is used for five different noisy speech signals. Further, the same signals are verified with deep neural network (DNN) model. In both the models, four hidden layers are used to analyse the noisy signal. However, for ADALINE case, the learning method used is weights and bias, whereas the restricted Boltzmann machines (RBMs) learning algorithm is used in DNN. Perceptual evaluation of speech quality (PESQ) and signal-to-noise-ratio (SNR) parameters are considered for verification and comparison purpose. The DNN is found better in terms of verified parameters as compared to ADALINE model. Nevertheless, the ADALINE model can be omitted in such comparison to prove the adaptability for the developments of automated system.