Method of noise reduction of an audio signal

A method of processing an audio signal, for removing noise from an original noisy signal comprising a component of combined voice with a noise component, the noise component comprises itself a component of transient noise and component of pseudoestacionaria noise, characterized in that this method is a method of analysis of temporal coherence of the signal sampled noise comprising the steps of: a) determining a reference signal for application to the signal to noise processing (10, 18) itself to attenuate more importantly the voice components that noise components of the noisy signal, said processing comprising: a1) applying an adaptive algorithm linear prediction operating on a linear combination of previous samples of the noisy signal, and a2) determining said reference signal by subtraction, with compensation gap between the original signal with noise, not previously filtered and the signal delivered by the linear prediction algorithm; b) determination (24) of a probability of presence / absence of a priori speech from respective energy levels in the spectral domain of the noisy signal and the reference signal; c) use of this probability a priori speech absence for estimating a noise spectrum and deriving 25 (26) of the estimated noise signal with denoising speech signal.