AdaBoost Noise Estimator for Subspace based Speech Enhancement

Abstract

The presence of noise degrades quality and intelligibility of speech signal. Signal subspace is a technique to separate uncorrelated additive noise from the speech signal by decompose the signal into noise-only subspace and signal plus noise subspace. Unfortunately, this decomposition is effective only for white noise. For color noise as most of speech enhancement cases, noise estimation is mandatory. This paper investigates the performance of AdaBoost algorithm to update noise estimation for subspace-based speech enhancement. AdaBoost estimator classify a signal frame into two class, speech and non-speech. From time frame which is identified as non-speech, the estimator calculates the power spectrum and assumes it as noise power spectra. During the presence of speech, the noise is assumed equal to the last updated value. The simulation shows us that for many cases, AdaBoost-based estimator has better performance than continuous noise update such as connected time-frequency speech presence region.