An audio processing device comprises a) at least one input unit for providing a time-frequency representation Y(k,n) of an electric input signal representing sound consisting of target speech and noise signal components, where k and n are frequency band and time frame indices, respectively, b) a noise reduction system configured to b1) determine a first signal to noise ratio estimate γ(k,n) of said electric input signal, and to b2) determine a second signal to noise signal ratio estimate ζ(k,n) of said electric input signal from said first signal to noise ratio estimate γ(k,n) based on a recursive algorithm providing non-linear smoothing, and wherein a determination of said one or more bias and/or smoothing parameters comprises the use of supervised learning, e.g. one or more neural networks. The invention may be used in audio processing devices, such as hearing aids, headsets, ear phones, active ear protection systems, handsfree telephone systems, mobile telephones, etc.