Speech noise reduction using adaptive spline neural networks

A new Neural Network architecture for real-time oriented speech denoising is proposed. It is based on Adaptive Spline neurons, whose peculiarity is the adaptive activation function. So, in the training phase, we can update both values of weights and activation function shape, obtaining networks with more flexibility and generalization capabilities. Net training is performed through the classical back-propagation rule. We focused our attention to continuous uncorrelated disturbs and we tried two kinds of approach: in the first one we processed the whole signal by a single network, while in the second one we operated a frequency sub-bands decomposition and we processed every sub-channel separately in a parallel way. The first approach is less heavy but the second one gives better results, due to the fact that, in practical application, background noise is frequency dependent. Results show improvements of Signal to Noise Ratio (SNR) and better performances in comparison with classical denoising neural networks.