For enhancing noisy signals, machine-learning based single-channel speechenhancement schemes exploit prior knowledge about typical speech spectralstructures. To ensure a good generalization and to meet requirements in termsof computational complexity and memory consumption, certain methods restrictthemselves to learning speech spectral envelopes. We refer to these approachesas machine-learning spectral envelope (MLSE)-based approaches. In this paper we show by means of theoretical and experimental analyses thatfor MLSE-based approaches, super-Gaussian priors allow for a reduction of noisebetween speech spectral harmonics which is not achievable using Gaussianestimators such as the Wiener filter. For the evaluation, we use a deep neuralnetwork (DNN)-based phoneme classifier and a low-rank nonnegative matrixfactorization (NMF) framework as examples of MLSE-based approaches. A listeningexperiment and instrumental measures confirm that while super-Gaussian priorsyield only moderate improvements for classic enhancement schemes, forMLSE-based approaches super-Gaussian priors clearly make an importantdifference and significantly outperform Gaussian priors.
展开▼