Unscented Transform-Based Dual-Channel Noise Estimation: Application to Speech Enhancement on Smartphones

Some speech processing approaches rely on a noise estimation stage and their performance depends on the accuracy of the noise estimates. In this paper we propose a novel minimum mean square error (MMSE) noise estimator that takes advantage of dual-channel noisy observations to avoid the use of a clean speech model while keeping a simple formulation. The parameters of this estimator are obtained through the unscented transform (UT), which is able to compute better quality statistics in a more efficient way than through classical vector Taylor series (VTS) linearization. For evaluation, speech enhancement on a dual-microphone smartphone in close-talk conditions is considered, which is a particular application of interest. Results show the superiority of our proposal with respect to other single- and dual-channel noise estimation methods in terms of different measures such as estimation accuracy as well as on the enhanced speech signal, i.e., speech quality and intelligibility.