Speech and Audio Enhancement Algorithms using Discriminative Random Fields

Abstract

We address the problem of single channel speech enhancement in additive noise. The ideal binary mask approach is considered. In this framework, we estimate the binary mask by classifying each time-frequency (TF) bin of the noisy signal as noise or speech. We use the Discriminative Random Fields (DRF) as the framework. The DRF based formulation comprises of two terms (i) an enhancement term (ii) a smoothing term. A likelihood ratio test based enhancement function is proposed for detecting speech or noise presence in each TF bin. Ising model is considered for smoothing to ensure temporal and spectral continuity in the estimated binary mask. Maximizing the joint distribution of noisy observations and latent labels over successive iterations results in reduced isolated TF components which in turn reduces musical noise. The Iterated Conditional Modes (ICM) algorithm is used to infer binary mask from the noisy signal. Thus, we estimate a 2-dimensional (2-D) ideal binary mask in spectro-temporal domain using the above framework, resulting in each TF bin being classified as noise or speech