Efficient speech de-noising applied to colored noise based dynamic low-pass filter supervised by cascade neural networks

In this paper, we investigated the enhancement of speech by applying an optimal adaptive low-pass filter supervised by neural network. The corruption of speech due to the presence of additive noise causes its degradation in quality and intelligibility. To filter this distorted signal in its spatial representation is a hard task. This task is more difficult to realize if the distortion are caused by colored noise. In addition using a static filter is not efficient due to the speech signal variability. In the same sentence a phoneme can change in shape and amplitude. For these constraints, we propose to apply a low-pass filter with Gaussian core supervised by neural networks. Filtering strength changes continuously with the phoneme variation to generate a variable filter that change over the whole sentence.