Wavelet-based noise estimation techniques for speech enhancement

In this paper, we describe the implementation of three noise estimation algorithms using two different wavelet decomposition methods: Second-generation and Perceptual wavelet packet transform. The three-presented algorithms are: (a) smoothing based adaptive noise estimation, (b) quantile based noise estimation and (c) minimum variance tracking-based noise estimation These algorithms, which do not need a speech activity detector nor signal statistics learning histograms, are based on estimating the noise power from the noisy speech itself. The performance of presented algorithms has been evaluated and compared for different noise types and levels. A new robust noise estimation technique utilizing a combination of the quantile-based and smoothing based algorithms has been proposed. Reported results demonstrate how these algorithms are capable to track different noise types adequately but with varying degree of accuracy.