Systems and methods for restoring speech components

A process for the recovery of distorted speech components of an audio signal which are distorted by noise reduction or noise cancellation comprises determining frequency ranges distorted and non-distorted frequency ranges in the audio signal. The distorted frequency ranges include ranges of the audio signal in which a speech distortion is present. It will be carried out using a model for the treatment of predictions of the audio signal into frequency ranges distorted iterations. The model is adapted to modify the audio signal and may comprise a deep neural network which is trained using spectral envelope of clear or undamaged audio signals. Prior to each iteration, the audio signal is restored in the non-distorted frequency ranges of values ​​of the audio signal prior to the first iteration; wherein the audio signal is processed in the distorted frequency ranges from zero in the first iteration. Iterations are terminated when discrepancies of the audio signal in a non-distorted frequency ranges meet predefined criteria.