Abstract:

Over the years, countless algorithms have been proposed to solve the problem of speech enhancement from a noisy mixture. Many have succeeded in improving at least parts of the signal, while often deteriorating others. Based on the assumption that different algorithms are likely to enjoy different qualities and suffer from different flaws, we investigate the possibility of combining the strengths of multiple speech enhancement algorithms, formulating the problem in an ensemble learning framework. As a first example of such a system, we consider the prediction of a time-frequency mask obtained from the clean speech, based on the outputs of various algorithms applied on the noisy mixture. We consider several approaches involving various notions of context and various machine learning algorithms for classification, in the case of binary masks, and regression, in the case of continuous masks. We show that combining several algorithms in this way can lead to an improvement in enhancement performance, while simple averaging or voting techniques fail to do so.