From non parametric statistics to speech denoising

Dominique Pastor (1) , Asmaa Amehraye (1)

1 SC - Département Signal et Communications

Given some signal additively corrupted by independent white Gaussian noise with unknown standard deviation σ,we present a new estimator ofσ. This estimator derives from a theoretical result presented and commented in the paper. Without any preliminary signal detection, the esti-mate is performed on the basis of the time-frequency components returned by a standard spectrogram where the Discrete Fourier Transform is simply weighted by the square window. No assumption about the signal statistics is made.The signal time-frequency components are assumed to have probabilities of presence less than or equal to one half.This estimator is suited to speech denoising. It avoids the use of any Voice Activity Detector and is an alternative solution to subspace approaches. Objective performance measurements show that the standard Wiener filtering of speech signals can be tuned with the outcome of this es-timator without a significant loss in comparison with the measurements obtained when the noise standard deviation is known

Mots clés: Non parametric statistics; Speech denoising

Domaines: Traitement du signal et de l'image [eess.SP]

Dominique Pastor, Asmaa Amehraye. From non parametric statistics to speech denoising. ISIVC 2006 : 3d international symposium on image/video communications over fixed and mobile networks, Sep 2006, Hammamet, Tunisia. ⟨hal-02136901⟩